exo v1

.clauderules

@@ -0,0 +1,63 @@
+# Claude Code Rules - Follow Every Rule Exactly
+
+You must prioritize straightforward code semantics, well-named types, clear function signatures, and robust, carefully-chosen abstractions. Think about how your decisions might impact these aspects of code quality before proposing any changes.
+
+You have access to all modern Python features from Python 3.13, 3.12, 3.11...
+
+**When you're done making changes, remove any redundant comments; remaining comments should only apply to complex code segments, adding relevant context.**
+
+## 1. Code Discipline
+
+* Eliminate superfluous `try`/`catch` and `if` branches through strict typing and static analysis.
+* Use pure functions unless you must mutate fixed state—then wrap that state in a class.
+* Every function is **referentially transparent**: same inputs ⇒ same outputs, no hidden state, no unintended I/O.
+* Put side-effects in injectable "effect handlers"; keep core logic pure.
+
+## 2. Naming
+
+* Choose descriptive, non-abbreviated names—no 3-letter acronyms or non-standard contractions.
+* Anyone reading a function's type signature alone should grasp its purpose without extra context.
+
+## 3. Typing
+
+* Maintain **strict, exhaustive** typing; never bypass the type-checker.
+* Default to `Literal[...]` when an enum-like set is needed.
+* Prefer built-in types; when two values share structure but differ in meaning, enforce separation:
+  * Use `typing.NewType` for primitives (zero runtime cost).
+  * For serializable objects, add a `type: str` field that states the object's identity.
+
+## 4. Pydantic
+
+* Read, respect, and rely on Pydantic documentation.
+* Centralize a common `ConfigDict` with `frozen=True` and `strict=True` (or stricter) and reuse it everywhere.
+* For hierarchies of `BaseModel` variants, declare a discriminated union with `typing.Annotated[Base, Field(discriminator='variant')]`; publish a single `TypeAdapter[Base]` so all variants share one strict validator.
+
+## 5. IDs & UUIDs
+
+* Subclass Pydantic's `UUID4` for custom ID types.
+* Generate fresh IDs with `uuid.uuid4()`.
+* Create idempotency keys by hashing *persisted* state plus a **function-specific salt** to avoid collisions after crashes.
+
+## 6. Error Handling
+
+* Catch an exception **only** where you can handle or transform it meaningfully.
+* State in the docstring **where** each exception is expected to be handled and **why**.
+
+## 7. Dependencies
+
+* Introduce new external dependencies only after approval.
+* Request only libraries common in production environments.
+
+## 8. Use of `@final` & Freezing
+
+* Mark classes, methods, and variables as `@final` or otherwise immutable wherever applicable.
+
+## 9. Repository Workflow
+
+If you spot a rule violation within code that you've not been asked to work on directly, inform the user rather than patching it ad-hoc.
+
+---
+
+### One-Sentence Summary
+
+Write strictly-typed, pure, self-describing Python that uses Pydantic, well-scoped side-effects, immutable state, approved dependencies, and explicit error handling.

.cursorrules

@@ -0,0 +1,64 @@
+# follow **every** rule exactly; report any violation instead of silently fixing it.
+
+You must prioritize straightforward code semantics, well-named types, clear function signatures, and robust, carefully-chosen abstractions. Think about how your decisions might impact these aspects of code quality before proposing any changes.
+
+You can use the advanced features of `typing`. You have access to all of the new features from Python 3.13, 3.12, 3.11...
+
+**When you're done making your changes, remove any redundant comments that you may have left; the comments that remain should only apply to complex segments of code, adding relevant context.**
+
+## 1. Code Discipline
+
+* Eliminate superfluous `try` / `catch` and `if` branches through strict typing and static analysis.
+* Use pure functions unless you must mutate fixed state—then wrap that state in a class.
+* Every function is **referentially transparent**: same inputs ⇒ same outputs, no hidden state, no unintended I/O.
+* Put side-effects in injectable “effect handlers”; keep core logic pure.
+
+## 2. Naming
+
+* Choose descriptive, non-abbreviated names—no 3-letter acronyms or non-standard contractions.
+* Anyone reading a function’s type signature alone should grasp its purpose without extra context.
+
+## 3. Typing
+
+* Maintain **strict, exhaustive** typing; never bypass the type-checker.
+* Default to `Literal[...]` when an enum-like set is needed.
+* Prefer built-in types; when two values share structure but differ in meaning, enforce separation:
+  * Use `typing.NewType` for primitives (zero runtime cost).
+  * For serialisable objects, add a `type: str` field that states the object’s identity.
+
+## 4. Pydantic
+
+* Read, respect, and rely on Pydantic docs.
+* Centralise a common `ConfigDict` with `frozen=True` and `strict=True` (or stricter) and reuse it everywhere.
+* For hierarchies of `BaseModel` variants, declare a discriminated union with `typing.Annotated[Base, Field(discriminator='variant')]`; publish a single `TypeAdapter[Base]` so all variants share one strict validator.
+
+## 5. IDs & UUIDs
+
+* Subclass Pydantic’s `UUID4` for custom ID types.
+* Generate fresh IDs with `uuid.uuid4()`.
+* Create idempotency keys by hashing *persisted* state plus a **function-specific salt** to avoid collisions after crashes.
+
+## 6. Error Handling
+
+* Catch an exception **only** where you can handle or transform it meaningfully.
+* State in the docstring **where** each exception is expected to be handled and **why**.
+
+## 7. Dependencies
+
+* Introduce new external dependencies only after approval.
+* Request only libraries common in production environments.
+
+## 8. Use of `@final` & Freezing
+
+* Mark classes, methods, and variables as `@final` or otherwise immutable wherever applicable.
+
+## 9. Repository Workflow
+
+If you spot a rule violation within code that you've not been asked to work on directly, inform the user rather than patching it ad-hoc.
+
+
+---
+
+### One-Sentence Summary
+
+Write strictly-typed, pure, self-describing Python that uses Pydantic, well-scoped side-effects, immutable state, approved dependencies, and explicit error handling

.envrc

@@ -0,0 +1 @@
+use flake

.githooks/post-checkout

@@ -0,0 +1,3 @@
+#!/bin/sh
+command -v git-lfs >/dev/null 2>&1 || { printf >&2 "\n%s\n\n" "This repository is configured for Git LFS but 'git-lfs' was not found on your path. If you no longer wish to use Git LFS, remove this hook by deleting the 'post-checkout' file in the hooks directory (set by 'core.hookspath'; usually '.git/hooks')."; exit 2; }
+git lfs post-checkout "$@"

.githooks/post-commit

@@ -0,0 +1,3 @@
+#!/bin/sh
+command -v git-lfs >/dev/null 2>&1 || { printf >&2 "\n%s\n\n" "This repository is configured for Git LFS but 'git-lfs' was not found on your path. If you no longer wish to use Git LFS, remove this hook by deleting the 'post-commit' file in the hooks directory (set by 'core.hookspath'; usually '.git/hooks')."; exit 2; }
+git lfs post-commit "$@"

.githooks/post-merge

@@ -0,0 +1,3 @@
+#!/bin/sh
+command -v git-lfs >/dev/null 2>&1 || { printf >&2 "\n%s\n\n" "This repository is configured for Git LFS but 'git-lfs' was not found on your path. If you no longer wish to use Git LFS, remove this hook by deleting the 'post-merge' file in the hooks directory (set by 'core.hookspath'; usually '.git/hooks')."; exit 2; }
+git lfs post-merge "$@"

.githooks/pre-push

@@ -0,0 +1,3 @@
+#!/bin/sh
+command -v git-lfs >/dev/null 2>&1 || { printf >&2 "\n%s\n\n" "This repository is configured for Git LFS but 'git-lfs' was not found on your path. If you no longer wish to use Git LFS, remove this hook by deleting the 'pre-push' file in the hooks directory (set by 'core.hookspath'; usually '.git/hooks')."; exit 2; }
+git lfs pre-push "$@"

.github/CODEOWNERS

@@ -0,0 +1,3 @@
+*   @ToxicPine
+*   @AlexCheema
+*   @GeluVrabie

.github/actions/conditional-commit/action.yml

@@ -0,0 +1,16 @@
+name: Commit if changed
+description: "Create a commit when the working tree is dirty"
+
+inputs:
+  message:
+    description: "Commit message"
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Commit changed files
+      shell: bash
+      run: |
+        git diff --quiet && exit 0
+        git commit -am "${{ inputs.message }}"

.github/actions/format/action.yml

@@ -0,0 +1,10 @@
+name: Format Code
+
+description: "Run code formatter"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Format code
+      run: nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just fmt
+      shell: bash

.github/actions/lint-check/action.yml

@@ -0,0 +1,10 @@
+name: Lint Check
+
+description: "Check for lint errors"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Lint check
+      run: nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just lint-check
+      shell: bash

.github/actions/lint/action.yml

@@ -0,0 +1,10 @@
+name: Lint Code
+
+description: "Run code linter"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Lint code
+      run: nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just lint
+      shell: bash

.github/actions/regenerate-protobufs/action.yml

@@ -0,0 +1,10 @@
+name: Regenerate Protobufs
+
+description: "Regenerate protobuf files"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Regenerate protobufs
+      run: nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just regenerate-protobufs
+      shell: bash

.github/actions/setup-python-uv/action.yml

@@ -0,0 +1,20 @@
+name: Setup Python & uv
+
+description: "Regenerate Python environment from uv.lock"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Install uv
+      uses: astral-sh/setup-uv@v6
+      with:
+        enable-cache: true
+        cache-dependency-glob: uv.lock
+
+    - name: Install Python
+      run: uv python install
+      shell: bash
+
+    - name: Sync
+      run: uv sync --locked --all-extras --dev
+      shell: bash

.github/actions/typecheck/action.yml

@@ -0,0 +1,12 @@
+name: Type Check
+
+description: "Run type checker"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Run type checker
+      run: |
+        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just sync
+        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just check
+      shell: bash

.github/actions/unit-test/action.yml

@@ -0,0 +1,12 @@
+name: Unit Test
+
+description: "Run unit tests"
+
+runs:
+  using: "composite"
+  steps:
+    - name: Run unit tests
+      run: |
+        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just sync-clean
+        nix --extra-experimental-features nix-command --extra-experimental-features flakes develop -c just test-fast
+      shell: bash

.github/actions/verify-clean/action.yml

@@ -0,0 +1,20 @@
+name: Verify Clean Working Tree
+
+description: "Fail the job if the previous step left the working tree dirty"
+
+inputs:
+  step:
+    description: "The name of the step that just executed"
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Check git diff
+      shell: bash
+      run: |
+        if ! git diff --quiet; then
+          echo "Error: ${{ inputs.step }} left working tree dirty." >&2
+          git --no-pager diff >&2
+          exit 1
+        fi 

.github/benchmark-dashboard/README.md

@@ -0,0 +1,159 @@
+# EXO Benchmark Dashboard
+
+A fully self-contained, browser-based dashboard for tracking EXO benchmark performance over time.
+
+## Features
+
+- 📊 **Success Rate Tracking**: Monitor cluster reliability across commits
+- ⚡ **Response Time Analysis**: Track average request completion times  
+- 🎯 **Throughput Metrics**: Tokens per second visualization
+- 📈 **Request Distribution**: Success/failure breakdown over time
+- 🔄 **Auto-Refresh**: Updates every 60 seconds
+- 📺 **TV-Ready**: Large, clear visualizations perfect for display
+- 🔐 **Secure**: Credentials stored in browser localStorage only
+- 🌐 **No Backend**: Directly accesses S3 from the browser
+
+## Quick Start
+
+### Option 1: Direct File Access (Simplest)
+
+Just open the HTML file directly in your browser:
+
+```bash
+open .github/benchmark-dashboard/index.html
+```
+
+Then click "Configure AWS Credentials" and enter your keys.
+
+### Option 2: URL Parameters (For Quick Setup)
+
+```bash
+# Serve with credentials in URL (they'll be moved to localStorage)
+open ".github/benchmark-dashboard/index.html?accessKey=YOUR_KEY&secretKey=YOUR_SECRET&region=us-east-1"
+```
+
+The credentials will be saved to localStorage and removed from the URL immediately.
+
+### Option 3: Simple HTTP Server
+
+```bash
+# From repo root
+python3 -m http.server 8080
+
+# Then open: http://localhost:8080/.github/benchmark-dashboard/
+```
+
+## AWS Credentials
+
+The dashboard needs read-only access to the `exo-benchmark-results` S3 bucket.
+
+### Required IAM Permissions
+
+```json
+{
+  "Version": "2012-10-17",
+  "Statement": [
+    {
+      "Effect": "Allow",
+      "Action": [
+        "s3:GetObject",
+        "s3:ListBucket"
+      ],
+      "Resource": [
+        "arn:aws:s3:::exo-benchmark-results",
+        "arn:aws:s3:::exo-benchmark-results/*"
+      ]
+    }
+  ]
+}
+```
+
+### Security Notes
+
+- ✅ Credentials stored in browser `localStorage` only
+- ✅ Never sent to any server (except AWS)
+- ✅ All S3 access happens client-side
+- ✅ Use read-only IAM credentials
+- ⚠️ Don't commit credentials to git
+- ⚠️ Use a dedicated read-only IAM user
+
+## TV/Kiosk Mode
+
+For permanent display on a TV:
+
+### macOS
+```bash
+open -a "Google Chrome" --args --kiosk ".github/benchmark-dashboard/index.html"
+```
+
+### Linux
+```bash
+chromium-browser --kiosk --app="file://$(pwd)/.github/benchmark-dashboard/index.html"
+```
+
+### Auto-start on Boot
+
+Create a simple startup script:
+
+```bash
+#!/bin/bash
+# /usr/local/bin/start-benchmark-dashboard.sh
+
+cd /path/to/exo
+python3 -m http.server 8080 &
+sleep 2
+chromium-browser --kiosk http://localhost:8080/.github/benchmark-dashboard/
+```
+
+## Data Displayed
+
+### Summary Cards
+- **Latest Success Rate**: Most recent benchmark success percentage with trend
+- **Avg Response Time**: Latest average response time in ms with trend
+- **Total Benchmarks**: Count of all benchmarks run
+- **Active Configurations**: Number of unique benchmark configs
+
+### Charts
+1. **Success Rate Over Time**: Line chart showing reliability trends
+2. **Average Response Time**: Performance over time (lower is better)
+3. **Throughput**: Tokens/second metric (higher is better)
+4. **Request Distribution**: Stacked bar chart of successes/failures
+
+## How It Works
+
+1. **Loads AWS SDK**: Uses AWS SDK for JavaScript (browser version)
+2. **Lists S3 Objects**: Fetches all files from `s3://exo-benchmark-results/bench/`
+3. **Downloads Results**: Fetches each JSON result file
+4. **Parses & Visualizes**: Uses Chart.js to create interactive charts
+5. **Auto-Refreshes**: Polls S3 every 60 seconds for new results
+
+## Customization
+
+To modify the dashboard:
+
+1. Edit `index.html` 
+2. Adjust `REFRESH_INTERVAL` for different polling frequency
+3. Modify chart colors/styles in the Chart.js configuration
+4. Add new metrics by extending the results parsing
+
+## Troubleshooting
+
+**"AWS credentials not configured"**
+- Click "Configure AWS Credentials" and enter your keys
+
+**"Error loading benchmark data"**
+- Check AWS credentials are correct
+- Verify S3 bucket name is `exo-benchmark-results`
+- Ensure IAM user has read permissions
+- Check browser console for detailed errors
+
+**"No benchmark results found"**
+- Wait for benchmark workflows to run
+- Verify results are being uploaded to S3
+- Check S3 bucket has files in `bench/` prefix
+
+**Charts not updating**
+- Check browser console for errors
+- Verify network connectivity to S3
+- Try refreshing the page manually
+

.github/configs/README.md

@@ -0,0 +1,186 @@
+# EXO Benchmark Configurations
+
+This directory contains configuration files for the EXO staged benchmark system.
+
+## Overview
+
+The staged benchmark system allows you to run complex, multi-stage load tests against EXO clusters. Each stage can have different characteristics:
+
+- **Prompt Length**: Number of tokens in the input prompt
+- **Generation Length**: Maximum tokens to generate in the response
+- **Time Between Requests**: Delay (in seconds) between firing consecutive requests
+- **Iterations**: Number of requests to send in this stage
+
+Requests are **fire-and-forget** - they don't wait for the previous request to complete. This allows you to test overlapping request handling and measure success rates under load.
+
+## Configuration Files
+
+### `bench_simple.yaml`
+A minimal configuration that replicates the behavior of the original `bench.py` script:
+- Single stage with 1 iteration
+- Short prompt (~20 tokens)
+- Generates up to 100 tokens
+
+This is useful for quick smoke tests.
+
+### `bench_config.yaml`
+A comprehensive multi-stage benchmark with:
+1. **Warmup** (10 requests): Light load with short prompts
+2. **Medium Load** (20 requests): Moderate load with medium prompts
+3. **Stress Test** (30 requests): Heavy overlapping requests with long prompts
+4. **Cooldown** (5 requests): Light load to wind down
+
+This tests the cluster's behavior under varying load patterns.
+
+## Configuration Schema
+
+```yaml
+# Hardware configuration - maps runner labels to instance counts
+hardware_plan:
+  M3ULTRA_GPU80_512GB: 4
+
+# Environment variables to set on each node (optional)
+environment:
+  OVERRIDE_MEMORY_MB: 512
+
+# Timeout for instance and runner readiness (seconds)
+timeout_seconds: 600
+
+# Model instances to run concurrently
+model_ids:
+  - "mlx-community/Llama-3.2-1B-Instruct-4bit"
+
+# Benchmark stages
+stages:
+  - name: "stage_name"              # Human-readable name for this stage
+    prompt_length: 100               # Target prompt length in tokens
+    generation_length: 200           # Max tokens to generate
+    time_between_requests: 2.0       # Seconds between firing requests
+    iterations: 10                   # Number of requests in this stage
+```
+
+## Running Benchmarks
+
+### Via GitHub Actions
+
+**Automatic (every commit):**
+- The **`bench`** workflow runs automatically on every push
+- Uses `bench_simple.yaml` as the default configuration
+- All settings (hardware plan, timeout, environment variables, models, stages) are defined in the config file
+
+**Manual (on-demand):**
+1. Go to **Actions** → **bench** workflow
+2. Click **Run workflow**
+3. Configure:
+   - **Config File**: Path to your YAML config (default: `.github/configs/bench_simple.yaml`)
+     - `.github/configs/bench_simple.yaml` for quick tests
+     - `.github/configs/bench_config.yaml` for complex multi-stage tests
+   
+All other settings (hardware plan, timeout, environment variables, models, stages) are read from the specified config file.
+
+### Via Command Line
+
+```bash
+# Start EXO on localhost:8000
+uv run exo --api-port 8000
+
+# Run simple benchmark (1 stage, 1 iteration)
+python3 .github/scripts/bench.py \
+  --api-port 8000 \
+  --config .github/configs/bench_simple.yaml \
+  --expected-nodes 1 \
+  --is-primary true \
+  --timeout-seconds 600
+
+# Run complex staged benchmark (4 stages, multiple iterations)
+python3 .github/scripts/bench.py \
+  --api-port 8000 \
+  --config .github/configs/bench_config.yaml \
+  --expected-nodes 1 \
+  --is-primary true \
+  --timeout-seconds 600
+```
+
+## Output Metrics
+
+For each stage, the benchmark reports:
+
+- **Total Requests**: Number of requests fired
+- **Successful Requests**: Requests that completed successfully
+- **Failed Requests**: Requests that encountered errors
+- **Success Rate**: Percentage of successful requests
+- **Total Tokens**: Sum of all tokens generated across successful requests
+- **Avg Tokens/Request**: Average tokens per successful request
+- **Avg Time/Request**: Average completion time per successful request
+
+A JSON summary is also printed for easy parsing and storage.
+
+## Creating Custom Benchmarks
+
+To create a custom benchmark:
+
+1. Copy an existing config file (e.g., `bench_config.yaml`)
+2. Modify the stages to match your test scenario
+3. Save it in this directory with a descriptive name
+4. Run it using the workflow or command line
+
+### Example: Sustained Load Test
+
+```yaml
+hardware_plan:
+  M3ULTRA_GPU80_512GB: 2
+
+environment:
+  OVERRIDE_MEMORY_MB: 1024
+
+timeout_seconds: 600
+
+model_ids:
+  - "mlx-community/Llama-3.2-1B-Instruct-4bit"
+
+stages:
+  - name: "sustained_load"
+    prompt_length: 200
+    generation_length: 150
+    time_between_requests: 0.5     # Very fast - 2 requests/second
+    iterations: 100                 # Run for ~50 seconds
+```
+
+### Example: Varying Prompt Sizes
+
+```yaml
+hardware_plan:
+  M4PRO_GPU16_24GB: 3
+
+timeout_seconds: 900
+
+model_ids:
+  - "mlx-community/Llama-3.2-1B-Instruct-4bit"
+
+stages:
+  - name: "tiny_prompts"
+    prompt_length: 10
+    generation_length: 100
+    time_between_requests: 1.0
+    iterations: 10
+    
+  - name: "medium_prompts"
+    prompt_length: 200
+    generation_length: 100
+    time_between_requests: 1.0
+    iterations: 10
+    
+  - name: "large_prompts"
+    prompt_length: 1000
+    generation_length: 100
+    time_between_requests: 1.0
+    iterations: 10
+```
+
+## Tips
+
+- **Overlapping Requests**: Set `time_between_requests` < expected completion time to test concurrent request handling
+- **Sequential Requests**: Set `time_between_requests` > expected completion time to ensure requests don't overlap
+- **Realistic Load**: Model real usage patterns by varying prompt/generation lengths across stages
+- **Success Rate**: A 100% success rate indicates the cluster handled the load well; lower rates suggest capacity limits
+

.github/configs/bench_config.yaml

@@ -0,0 +1,49 @@
+# EXO Staged Benchmark Configuration
+# This configuration defines a multi-stage load test for EXO clusters
+
+# Hardware configuration - maps runner labels to instance counts
+hardware_plan:
+  M3ULTRA_GPU80_512GB: 4
+
+# Environment variables to set on each node (optional)
+environment:
+  OVERRIDE_MEMORY_MB: 512
+
+# Timeout for instance and runner readiness (seconds)
+timeout_seconds: 600
+
+# Multiple instances run concurrently on the cluster
+model_ids:
+  - "mlx-community/Qwen3-0.6B-4bit"
+  - "mlx-community/Qwen3-0.6B-4bit"
+
+# Stages run sequentially, each with its own characteristics
+stages:
+  # Stage 1: Light load with short prompts
+  - name: "warmup"
+    prompt_length: 50          # Number of tokens in prompt
+    generation_length: 100     # Max tokens to generate
+    time_between_requests: 5.0 # Seconds between firing requests
+    iterations: 10             # Number of requests to send in this stage
+    
+  # Stage 2: Medium load with medium prompts
+  - name: "medium_load"
+    prompt_length: 200
+    generation_length: 150
+    time_between_requests: 3.0
+    iterations: 20
+    
+  # Stage 3: Heavy load with long prompts - requests will overlap
+  - name: "stress_test"
+    prompt_length: 500
+    generation_length: 200
+    time_between_requests: 1.0  # Fast firing - will definitely overlap
+    iterations: 30
+    
+  # Stage 4: Cool down with simple prompts
+  - name: "cooldown"
+    prompt_length: 50
+    generation_length: 50
+    time_between_requests: 10.0
+    iterations: 5
+

.github/configs/bench_simple.yaml

@@ -0,0 +1,125 @@
+# Simple single-shot benchmark
+# Tests 2 instances concurrently on 2 nodes
+
+# Hardware configuration - maps runner labels to instance counts
+hardware_plan:
+  puffin4: 1
+  puffin8: 1
+
+# Environment variables to set on each node
+environment:
+  PLACEHOLDER: "placeholder"
+  # OVERRIDE_MEMORY_MB: 50000
+  MLX_METAL_FAST_SYNCH: 1
+
+# Timeout for instance and runner readiness (seconds)
+timeout_seconds: 1800
+
+# Model instances to run concurrently
+model_ids:
+  # - "mlx-community/DeepSeek-V3.1-8bit"
+  # - "mlx-community/Kimi-K2-Instruct-4bit"
+  - "mlx-community/Kimi-K2-Thinking"
+  # - "mlx-community/Qwen3-235B-A22B-4bit"
+  # - "mlx-community/Llama-3.3-70B-Instruct-4bit"
+  # - "mlx-community/Llama-3.3-70B-Instruct-8bit"
+  # - "mlx-community/Llama-3.2-1B-Instruct-4bit"
+
+# Sharding strategy: "Pipeline" or "Tensor"
+sharding: "Tensor"
+
+# Instance type: "MlxRing" or "MlxIbv"
+instance_meta: "MlxIbv"
+
+# If true, run requests sequentially (no overlap); if false, fire-and-forget (default: false)
+no_overlap: true
+
+# Benchmark stages
+# pp: 64, 256, 1024, 2048, 4096, 8192, 16384
+# g: 64, 512
+stages:
+  # - name: "simple"
+  #   prompt_length: 512
+  #   generation_length: 10
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp64_g64"
+  #   prompt_length: 64
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp64_g64"
+  #   prompt_length: 64
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp64_g512"
+  #   prompt_length: 64
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp256_g64"
+  #   prompt_length: 256
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  - name: "pp256_g64"
+    prompt_length: 256
+    generation_length: 64
+    time_between_requests: 2.0
+    iterations: 5
+  # - name: "pp256_g512"
+  #   prompt_length: 256
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp1024_g64"
+  #   prompt_length: 1024
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp1024_g512"
+  #   prompt_length: 1024
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp2048_g64"
+  #   prompt_length: 2048
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp2048_g512"
+  #   prompt_length: 2048
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp4096_g64"
+  #   prompt_length: 4096
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 4
+  # - name: "pp4096_g512"
+  #   prompt_length: 4096
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp8192_g64"
+  #   prompt_length: 8192
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp8192_g512"
+  #   prompt_length: 8192
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 5
+  # - name: "pp16384_g64"
+  #   prompt_length: 16384
+  #   generation_length: 64
+  #   time_between_requests: 2.0
+  #   iterations: 10
+  # - name: "pp16384_g512"
+  #   prompt_length: 16384
+  #   generation_length: 512
+  #   time_between_requests: 2.0
+  #   iterations: 10

.github/scripts/build_matrix.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+import json
+import os
+from typing import NotRequired, TypedDict, cast
+
+import yaml
+
+
+class MatrixEntry(TypedDict):
+    label: str
+    index: int
+
+
+class MatrixInclude(TypedDict):
+    label: str
+    index: int
+    is_primary: bool
+    expected_nodes: int
+
+
+class Config(TypedDict):
+    hardware_plan: dict[str, int]
+    timeout_seconds: NotRequired[int]
+    environment: NotRequired[dict[str, str]]
+
+
+# Read the config file
+config_file: str = os.environ["CONFIG_FILE"]
+with open(config_file, "r") as f:
+    config: Config = cast(Config, yaml.safe_load(f))
+
+# Extract hardware plan from config
+plan: dict[str, int] = config["hardware_plan"]
+if not plan:
+    raise ValueError(f"No hardware_plan found in {config_file}")
+
+# Build matrix entries
+entries: list[MatrixEntry] = []
+for label, count in plan.items():
+    for idx in range(count):
+        entries.append({"label": label, "index": idx})
+
+total_nodes: int = len(entries)
+matrix: dict[str, list[MatrixInclude]] = {
+    "include": [
+        {
+            "label": e["label"],
+            "index": e["index"],
+            "is_primary": (i == 0),
+            "expected_nodes": total_nodes,
+        }
+        for i, e in enumerate(entries)
+    ]
+}
+
+# Extract other config values
+timeout_seconds: int = config.get("timeout_seconds", 600)
+environment: dict[str, str] = config.get("environment", {})
+
+# Output to GitHub Actions
+with open(os.environ["GITHUB_OUTPUT"], "a") as f:
+    f.write(f"matrix={json.dumps(matrix)}\n")
+    f.write(f"config_file={config_file}\n")
+    f.write(f"timeout_seconds={timeout_seconds}\n")
+    f.write(f"environment={json.dumps(environment)}\n")
+
+print(f"Matrix: {json.dumps(matrix)}")
+print(f"Config file: {config_file}")
+print(f"Timeout: {timeout_seconds}")
+print(f"Environment: {json.dumps(environment)}")

.github/workflows/BENCH_USAGE.md

@@ -0,0 +1,156 @@
+# Benchmark Workflow Usage
+
+## Overview
+
+The `bench_matrix.yml` workflow enables distributed benchmarking of models across multiple self-hosted macOS runners with different hardware configurations.
+
+## Workflow Inputs
+
+| Input | Description | Default | Required |
+|-------|-------------|---------|----------|
+| `model_id` | Model ID to benchmark | `mlx-community/Llama-3.2-1B-Instruct-4bit` | Yes |
+| `hardware_plan` | JSON mapping of runner labels to counts | `{"M4PRO_GPU16_24GB": 1}` | Yes |
+| `prompt` | Benchmark prompt text | `What is the capital of France?` | No |
+| `timeout_seconds` | Timeout for instance/runner readiness | `600` | No |
+
+## Hardware Plan Format
+
+The `hardware_plan` input is a JSON object mapping runner labels to the number of machines:
+
+```json
+{
+  "M4PRO_GPU16_24GB": 2,
+  "M3ULTRA_GPU80_512GB": 1
+}
+```
+
+This example would:
+- Start 2 runners with the `M4PRO_GPU16_24GB` label
+- Start 1 runner with the `M3ULTRA_GPU80_512GB` label
+- Total of 3 runners coordinating on a single distributed inference instance
+
+## How It Works
+
+1. **Planning Job** (`plan`)
+   - Runs on `ubuntu-latest`
+   - Parses the `hardware_plan` JSON
+   - Generates a dynamic matrix with one entry per runner
+   - Only the first runner (index 0) is marked as `is_primary`
+
+2. **Benchmark Worker Jobs** (`bench_worker`)
+   - Each job runs on a self-hosted macOS runner with the specified label
+   - All runners start EXO in parallel
+   - The primary runner creates the model instance
+   - All runners wait for their assigned runner to be ready (Loaded/Running status)
+   - The primary runner executes the benchmark and prints results
+   - The primary runner deletes the instance
+
+## Example Usage
+
+### Single Machine Benchmark
+
+```yaml
+model_id: mlx-community/Llama-3.2-1B-Instruct-4bit
+hardware_plan: '{"M4PRO_GPU16_24GB": 1}'
+prompt: What is the capital of France?
+timeout_seconds: 600
+```
+
+### Multi-Machine Distributed Benchmark
+
+```yaml
+model_id: mlx-community/Llama-3.2-3B-Instruct-4bit
+hardware_plan: '{"M4PRO_GPU16_24GB": 2, "M3ULTRA_GPU80_512GB": 1}'
+prompt: Explain quantum computing in simple terms.
+timeout_seconds: 900
+```
+
+## Benchmark Output
+
+The primary runner outputs a JSON object with benchmark results:
+
+```json
+{
+  "model_id": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+  "instance_id": "abc-123-def",
+  "tokens": 42,
+  "elapsed_s": 2.451,
+  "tps": 17.136
+}
+```
+
+Where:
+- `tokens`: Number of chunks/tokens generated
+- `elapsed_s`: Total elapsed time in seconds
+- `tps`: Tokens per second (tokens / elapsed_s)
+
+## Runner Requirements
+
+Each self-hosted runner must:
+- Be labeled with appropriate hardware tags (e.g., `M4PRO_GPU16_24GB`)
+- Have the `self-hosted` and `macOS` labels
+- Have Nix installed with flakes enabled
+- Have network connectivity to other runners in the same job
+
+## Architecture
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│ GitHub Actions Workflow (bench_matrix.yml)                  │
+├─────────────────────────────────────────────────────────────┤
+│                                                              │
+│  ┌────────────────┐                                         │
+│  │  Plan Job      │                                         │
+│  │  (ubuntu)      │──┬─► Matrix: [{label, index, primary}] │
+│  └────────────────┘  │                                      │
+│                      │                                      │
+│  ┌───────────────────▼──────────────────────────────────┐  │
+│  │  Bench Worker Jobs (Matrix)                         │  │
+│  ├──────────────────────────────────────────────────────┤  │
+│  │                                                       │  │
+│  │  Runner 0 (Primary)     Runner 1         Runner 2    │  │
+│  │  ┌─────────────┐       ┌─────────────┐ ┌──────────┐ │  │
+│  │  │ Start EXO   │       │ Start EXO   │ │ Start EXO│ │  │
+│  │  │ Create Inst │       │ Wait...     │ │ Wait...  │ │  │
+│  │  │ Wait Ready  │       │ Wait Ready  │ │ Wait...  │ │  │
+│  │  │ Run Bench   │       │ (idle)      │ │ (idle)   │ │  │
+│  │  │ Print TPS   │       │             │ │          │ │  │
+│  │  │ Delete Inst │       │             │ │          │ │  │
+│  │  └─────────────┘       └─────────────┘ └──────────┘ │  │
+│  └───────────────────────────────────────────────────────┘  │
+└─────────────────────────────────────────────────────────────┘
+```
+
+## Implementation Details
+
+### `scripts/bench.py`
+
+A standalone Python script that:
+- Creates instance (primary only)
+- Polls `/state` endpoint until instance and all runners are ready
+- Executes chat completion with timing (primary only)
+- Parses SSE stream and counts tokens
+- Computes TPS metrics
+- Cleans up instance (primary only)
+
+### Key Functions
+
+- `wait_for_instance()`: Polls until instance with model_id appears
+- `wait_for_runners_ready()`: Polls until expected number of runners reach Loaded/Running status
+- `run_benchmark()`: Executes chat completion, measures time, counts tokens
+
+## Troubleshooting
+
+### Instance never becomes ready
+- Check EXO logs in the workflow output
+- Verify model_id is valid and accessible
+- Increase `timeout_seconds`
+
+### Runner mismatch
+- Ensure hardware_plan counts match available labeled runners
+- Check runner labels match exactly (case-sensitive)
+
+### Network issues
+- Verify runners can communicate on the network
+- Check firewall rules between runner hosts
+

.github/workflows/bench.yml

@@ -0,0 +1,305 @@
+name: bench
+
+on: [push]
+
+jobs:
+  plan:
+    if: contains(github.event.head_commit.message, '/bench')
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.build.outputs.matrix }}
+      config_file: ${{ steps.build.outputs.config_file }}
+      timeout_seconds: ${{ steps.build.outputs.timeout_seconds }}
+      environment: ${{ steps.build.outputs.environment }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Build matrix from config file
+        id: build
+        shell: bash
+        run: |
+          set -euo pipefail
+          CONFIG_FILE='.github/configs/bench_simple.yaml'
+          export CONFIG_FILE
+          echo "Config file: $CONFIG_FILE"
+          python3 .github/scripts/build_matrix.py
+
+  bench_worker:
+    needs: plan
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.plan.outputs.matrix) }}
+    name: "bench on ${{ matrix.label }} [${{ matrix.index }}]"
+    runs-on: [self-hosted, macOS, "${{ matrix.label }}"]
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          lfs: false
+
+      - name: Configure git user
+        run: |
+          git config --local user.email "github-actions@users.noreply.github.com"
+          git config --local user.name  "github-actions bot"
+        shell: bash
+
+      # TODO: this is mega hacky and I'd like a simpler solution.
+      - name: Setup Nix Environment
+        run: |
+          echo "Checking for nix installation..."
+          
+          # Check if nix is already available
+          if command -v nix >/dev/null 2>&1; then
+            echo "Nix already in PATH"
+          # Try sourcing profile scripts to set up environment properly
+          elif [ -f /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh ]; then
+            echo "Sourcing multi-user nix-daemon profile script"
+            source /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh
+          elif [ -f "$HOME/.nix-profile/etc/profile.d/nix.sh" ]; then
+            echo "Sourcing single-user nix profile script"
+            source "$HOME/.nix-profile/etc/profile.d/nix.sh"
+          elif [ -f /nix/var/nix/profiles/per-user/$USER/profile/etc/profile.d/nix.sh ]; then
+            echo "Sourcing per-user nix profile script"
+            source /nix/var/nix/profiles/per-user/$USER/profile/etc/profile.d/nix.sh
+          elif [ -f /etc/profile.d/nix.sh ]; then
+            echo "Sourcing system-wide nix profile script"
+            source /etc/profile.d/nix.sh
+          # Fallback: manually add nix to PATH if binary exists
+          elif [ -f /nix/var/nix/profiles/default/bin/nix ]; then
+            echo "Found nix binary, manually adding to PATH"
+            export PATH="/nix/var/nix/profiles/default/bin:$PATH"
+          elif [ -f "$HOME/.nix-profile/bin/nix" ]; then
+            echo "Found nix binary in user profile, manually adding to PATH"
+            export PATH="$HOME/.nix-profile/bin:$PATH"
+          else
+            echo "Nix not found. Debugging info:"
+            echo "USER: $USER"
+            echo "HOME: $HOME"
+            echo "Current PATH: $PATH"
+            echo ""
+            echo "Checking common Nix locations:"
+            echo "  /nix/var/nix/profiles/default/bin/nix:"
+            ls -la /nix/var/nix/profiles/default/bin/nix 2>/dev/null || echo "    Not found"
+            echo "  /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh:"
+            ls -la /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh 2>/dev/null || echo "    Not found"
+            echo "  ~/.nix-profile/etc/profile.d/nix.sh:"
+            ls -la "$HOME/.nix-profile/etc/profile.d/nix.sh" 2>/dev/null || echo "    Not found"
+            echo "  /nix/var/nix/profiles/per-user/$USER/profile/etc/profile.d/nix.sh:"
+            ls -la "/nix/var/nix/profiles/per-user/$USER/profile/etc/profile.d/nix.sh" 2>/dev/null || echo "    Not found"
+            echo ""
+            echo "/nix directory structure:"
+            ls -la /nix 2>/dev/null || echo "    /nix directory not found"
+            echo ""
+            echo "/nix/var:"
+            ls -la /nix/var 2>/dev/null || echo "    /nix/var not found"
+            echo ""
+            echo "/nix/store:"
+            ls -la /nix/store 2>/dev/null | head -20 || echo "    /nix/store not found"
+            echo ""
+            echo "GitHub Actions runner is running as user '$USER'."
+            echo "If Nix is installed for a different user, either:"
+            echo "  1. Install Nix for user '$USER' (multi-user install recommended)"
+            echo "  2. Configure the runner service to run as the user with Nix installed"
+            echo "  3. Ensure Nix is installed system-wide with proper daemon setup"
+            exit 1
+          fi
+          
+          # Verify nix is available and persist to GITHUB_ENV
+          if command -v nix >/dev/null 2>&1; then
+            echo "✓ Nix is available"
+            nix --version
+            echo "PATH=$PATH" >> $GITHUB_ENV
+            if [ -n "$NIX_PATH" ]; then
+              echo "NIX_PATH=$NIX_PATH" >> $GITHUB_ENV
+            fi
+          else
+            echo "ERROR: Failed to set up Nix"
+            echo "PATH after setup attempt: $PATH"
+            exit 1
+          fi
+        shell: bash
+
+      - name: Setup EXO_HOME and API_PORT
+        run: |
+          EXO_HOME=$(mktemp -d -t exo-e2e-XXXXXXXX)
+          API_PORT=$((49152 + RANDOM % (65535 - 49152 + 1)))
+          EXO_MODELS_DIR="$HOME/.exo/models"
+          EXO_LIBP2P_NAMESPACE="bench-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
+          echo "EXO_HOME=$EXO_HOME" >> "$GITHUB_ENV"
+          echo "API_PORT=$API_PORT" >> "$GITHUB_ENV"
+          echo "EXO_MODELS_DIR=$EXO_MODELS_DIR" >> "$GITHUB_ENV"
+          echo "EXO_LIBP2P_NAMESPACE=$EXO_LIBP2P_NAMESPACE" >> "$GITHUB_ENV"
+          echo "Created EXO_HOME: $EXO_HOME"
+          echo "Generated API_PORT: $API_PORT"
+          echo "Using models from: $EXO_MODELS_DIR"
+          echo "Using libp2p namespace: $EXO_LIBP2P_NAMESPACE"
+        shell: bash
+
+      - name: Configure local MLX if available
+        run: |
+          echo "=== DEBUG: Checking for local MLX configuration ==="
+          MODIFIED=false
+          
+          echo "Checking for /Users/Shared/mlx directory..."
+          if [ -d "/Users/Shared/mlx" ]; then
+            echo "✓ Found /Users/Shared/mlx"
+            ls -la /Users/Shared/mlx | head -5
+            echo "Enabling local mlx path in pyproject.toml"
+            sed -i.bak 's|^# mlx = { path = "/Users/Shared/mlx", editable=true }$|mlx = { path = "/Users/Shared/mlx", editable=true }|' pyproject.toml
+            MODIFIED=true
+          else
+            echo "✗ /Users/Shared/mlx not found, will use PyPI version"
+          fi
+          
+          echo "Checking for /Users/Shared/mlx-lm directory..."
+          if [ -d "/Users/Shared/mlx-lm" ]; then
+            echo "✓ Found /Users/Shared/mlx-lm"
+            ls -la /Users/Shared/mlx-lm | head -5
+            echo "Enabling local mlx-lm path in pyproject.toml"
+            sed -i.bak 's|^# mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }$|mlx-lm = { path = "/Users/Shared/mlx-lm", editable=true }|' pyproject.toml
+            MODIFIED=true
+          else
+            echo "✗ /Users/Shared/mlx-lm not found, will use PyPI version"
+          fi
+          
+          if [ "$MODIFIED" = true ]; then
+            echo "=== Modified pyproject.toml [tool.uv.sources] section: ==="
+            sed -n '/\[tool\.uv\.sources\]/,/^\[/{/^\[tool\.uv\.sources\]/p; /^\[/!p;}' pyproject.toml
+            echo "=== Regenerating uv.lock with local MLX paths... ==="
+            nix --extra-experimental-features nix-command --extra-experimental-features flakes develop --command uv lock --upgrade-package mlx --upgrade-package mlx-lm
+            echo "✓ Lock file regenerated"
+          else
+            echo "⚠ No local MLX directories found, using PyPI packages"
+          fi
+          echo "=== DEBUG: Local MLX configuration complete ==="
+        shell: bash
+
+      - name: Sync dependencies
+        run: |
+          if [ -d "/Users/Shared/test" ]; then
+            pushd /Users/Shared/test
+            uv sync --reinstall
+            popd
+          fi
+          echo "Running just sync to ensure clean dependencies..."
+          nix --extra-experimental-features nix-command --extra-experimental-features flakes develop --command just sync
+        shell: bash
+
+      - name: Start EXO and run bench script
+        shell: bash
+        env:
+          IS_PRIMARY: ${{ matrix.is_primary }}
+          EXPECTED_NODES: ${{ matrix.expected_nodes }}
+          HARDWARE_LABEL: ${{ matrix.label }}
+          CONFIG_FILE: ${{ needs.plan.outputs.config_file }}
+          TIMEOUT_SECONDS: ${{ needs.plan.outputs.timeout_seconds }}
+          ENVIRONMENT_JSON: ${{ needs.plan.outputs.environment }}
+        run: |
+          set -euo pipefail
+
+          # Parse environment variables from config
+          ENV_VARS=""
+          if [ -n "$ENVIRONMENT_JSON" ] && [ "$ENVIRONMENT_JSON" != "{}" ]; then
+            ENV_VARS=$(echo "$ENVIRONMENT_JSON" | python3 -c "import sys, json; env = json.load(sys.stdin); print(' '.join([f'{k}={v}' for k, v in env.items()]))")
+          fi
+
+          echo "Starting EXO with API_PORT=${API_PORT} EXO_HOME=${EXO_HOME} EXO_LIBP2P_NAMESPACE=${EXO_LIBP2P_NAMESPACE}"
+          echo "Environment variables from config: $ENV_VARS"
+          LOG_FILE=/tmp/exo.log
+          : > "$LOG_FILE"
+
+          MASTER_FLAG=""
+          if [ "$IS_PRIMARY" = "true" ]; then
+            MASTER_FLAG="-m"
+          fi
+
+          nix --extra-experimental-features nix-command --extra-experimental-features flakes develop --command bash -c \
+            "EXO_HOME=$EXO_HOME EXO_MODELS_DIR=$EXO_MODELS_DIR EXO_LIBP2P_NAMESPACE=$EXO_LIBP2P_NAMESPACE $ENV_VARS PYTHONUNBUFFERED=1 PYTHONDEBUG=1 PYTHONPATH=. uv run exo $MASTER_FLAG --api-port $API_PORT" \
+            >> "$LOG_FILE" 2>&1 &
+
+          EXO_PID=$!
+          echo "Started EXO in background with PID: $EXO_PID"
+          echo "Log file: $LOG_FILE"
+
+          cleanup() {
+            echo '=== EXO log (tail) ==='
+            tail -n 300 "$LOG_FILE" || true
+            if ps -p "$EXO_PID" >/dev/null 2>&1; then
+              echo "Killing EXO (PID $EXO_PID)"
+              kill "$EXO_PID" || true
+            fi
+          }
+          trap cleanup EXIT
+
+          for i in $(seq 1 60); do
+            if curl -s "http://localhost:${API_PORT}/state" >/dev/null 2>&1; then
+              echo "EXO API ready"
+              break
+            fi
+            if ! ps -p "$EXO_PID" >/dev/null 2>&1; then
+              echo "EXO terminated early"; sed -n '1,200p' "$LOG_FILE" || true; exit 1
+            fi
+            sleep 1
+          done
+
+          RESULTS_FILE="/tmp/bench_results_${GITHUB_RUN_ID}_${GITHUB_RUN_ATTEMPT}_$(date +%s).json"
+          echo "Results will be saved to: $RESULTS_FILE"
+          echo "RESULTS_FILE=$RESULTS_FILE" >> "$GITHUB_ENV"
+
+          echo "Running bench script with config: $CONFIG_FILE, timeout: $TIMEOUT_SECONDS"
+          nix --extra-experimental-features nix-command --extra-experimental-features flakes develop --command bash -c \
+            "PYTHONUNBUFFERED=1 uv run --no-project --with pyyaml --with pydantic python .github/scripts/bench.py \
+              --api-port $API_PORT \
+              --config $CONFIG_FILE \
+              --expected-nodes ${EXPECTED_NODES} \
+              --is-primary ${IS_PRIMARY} \
+              --timeout-seconds ${TIMEOUT_SECONDS} \
+              --output $RESULTS_FILE \
+              --git-commit ${GITHUB_SHA} \
+              --hardware-labels ${HARDWARE_LABEL}"
+
+      - name: Install AWS CLI
+        if: always() && env.RESULTS_FILE && matrix.is_primary
+        run: |
+          if ! command -v aws &> /dev/null; then
+            echo "AWS CLI not found, installing..."
+            brew install awscli
+          else
+            echo "AWS CLI already installed"
+          fi
+        shell: bash
+
+      - name: Upload results to S3
+        if: always() && env.RESULTS_FILE && matrix.is_primary
+        env:
+          AWS_ACCESS_KEY_ID: ${{ secrets.S3_BENCHMARKS_AWS_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.S3_BENCHMARKS_AWS_SECRET_ACCESS_KEY }}
+          AWS_DEFAULT_REGION: us-east-1
+        run: |
+          echo "Checking for results file: $RESULTS_FILE"
+          echo "Is primary: ${{ matrix.is_primary }}"
+
+          if [ -f "$RESULTS_FILE" ]; then
+            TIMESTAMP=$(date -u +%Y/%m/%d/%H%M%S)
+            S3_KEY="bench/${TIMESTAMP}_${GITHUB_SHA:0:8}_${GITHUB_RUN_ID}.json"
+            echo "Uploading results to s3://exo-benchmark-results/$S3_KEY"
+
+            aws s3 cp "$RESULTS_FILE" "s3://exo-benchmark-results/$S3_KEY" \
+              --content-type application/json \
+              --metadata "commit=${GITHUB_SHA},run_id=${GITHUB_RUN_ID},branch=${GITHUB_REF_NAME}"
+
+            echo "Results uploaded successfully"
+            echo "View at: https://exo-benchmark-results.s3.amazonaws.com/$S3_KEY"
+          else
+            echo "Results file not found at: $RESULTS_FILE"
+            echo "Skipping upload"
+          fi
+        shell: bash
+
+      - name: Cleanup EXO_HOME
+        run: |
+          echo "Cleaning up EXO_HOME: $EXO_HOME"
+          rm -rf "$EXO_HOME"
+        shell: bash
+        if: always()

.github/workflows/pipeline.yml

@@ -0,0 +1,183 @@
+name: ci-pipeline
+
+on:
+  push:
+  pull_request:
+    branches:
+      - staging
+      - main
+
+jobs:
+  typecheck:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          lfs: false
+
+      - uses: cachix/install-nix-action@v31
+        with:
+          nix_path: nixpkgs=channel:nixos-unstable
+
+      - name: Configure git user
+        run: |
+          git config --local user.email "github-actions@users.noreply.github.com"
+          git config --local user.name  "github-actions bot"
+        shell: bash
+
+      - name: Pull LFS files
+        run: |
+          echo "Pulling Git LFS files..."
+          git lfs pull
+        shell: bash
+
+      - name: Setup Nix Environment
+        run: |
+          echo "Checking for nix installation..."
+          
+          # Check if nix binary exists directly
+          if [ -f /nix/var/nix/profiles/default/bin/nix ]; then
+            echo "Found nix binary at /nix/var/nix/profiles/default/bin/nix"
+            export PATH="/nix/var/nix/profiles/default/bin:$PATH"
+            echo "PATH=$PATH" >> $GITHUB_ENV
+            nix --version
+          elif [ -f /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh ]; then
+            echo "Found nix profile script, sourcing..."
+            source /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh
+            nix --version
+          elif command -v nix >/dev/null 2>&1; then
+            echo "Nix already in PATH"
+            nix --version
+          else
+            echo "Nix not found. Debugging info:"
+            echo "Contents of /nix/var/nix/profiles/default/:"
+            ls -la /nix/var/nix/profiles/default/ 2>/dev/null || echo "Directory not found"
+            echo "Contents of /nix/var/nix/profiles/default/bin/:"
+            ls -la /nix/var/nix/profiles/default/bin/ 2>/dev/null || echo "Directory not found"
+            exit 1
+          fi
+        shell: bash
+
+      - name: Configure basedpyright include for local MLX
+        run: |
+          RUNNER_LABELS='${{ toJSON(runner.labels) }}'
+          if echo "$RUNNER_LABELS" | grep -q "local_mlx"; then
+            if [ -d "/Users/Shared/mlx" ]; then
+              echo "Updating [tool.basedpyright].include to use /Users/Shared/mlx"
+              awk '
+                BEGIN { in=0 }
+                /^\[tool\.basedpyright\]/ { in=1; print; next }
+                in && /^\[/ { in=0 }  # next section
+                in && /^[ \t]*include[ \t]*=/ {
+                  print "include = [\"/Users/Shared/mlx\"]"
+                  next
+                }
+                { print }
+              ' pyproject.toml > pyproject.toml.tmp && mv pyproject.toml.tmp pyproject.toml
+
+              echo "New [tool.basedpyright] section:"
+              sed -n '/^\[tool\.basedpyright\]/,/^\[/p' pyproject.toml | sed '$d' || true
+            else
+              echo "local_mlx tag present but /Users/Shared/mlx not found; leaving pyproject unchanged."
+            fi
+          else
+            echo "Runner does not have 'local_mlx' tag; leaving pyproject unchanged."
+          fi
+        shell: bash
+
+      - uses: ./.github/actions/typecheck
+
+  nix-flake-check:
+    name: Check Nix flake
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          lfs: false
+
+      - uses: cachix/install-nix-action@v31
+        with:
+          nix_path: nixpkgs=channel:nixos-unstable
+
+      - name: Run nix flake check
+        run: |
+          nix flake check
+        shell: bash
+
+#  ci:
+#    needs: typecheck
+#    runs-on: ubuntu-latest
+#    permissions:
+#      contents: read
+#    env:
+#      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+#    steps:
+#      - name: Checkout repository
+#        uses: actions/checkout@v4
+#        with:
+#          fetch-depth: 0
+#          token: ${{ secrets.GITHUB_TOKEN }}
+#          lfs: true
+#
+#      - name: Configure git user
+#        run: |
+#          git config --local user.email "github-actions@users.noreply.github.com"
+#          git config --local user.name  "github-actions bot"
+#        shell: bash
+#
+#      - name: Pull LFS files
+#        run: |
+#          echo "Pulling Git LFS files..."
+#          git lfs pull
+#        shell: bash
+#
+#      - name: Setup EXO_HOME and API_PORT
+#        run: |
+#          EXO_HOME=$(mktemp -d -t exo-ci-XXXXXXXX)
+#          # Generate random port (macOS compatible method)
+#          API_PORT=$((49152 + RANDOM % (65535 - 49152 + 1)))
+#          echo "EXO_HOME=$EXO_HOME" >> $GITHUB_ENV
+#          echo "API_PORT=$API_PORT" >> $GITHUB_ENV
+#          echo "Created EXO_HOME: $EXO_HOME"
+#          echo "Generated API_PORT: $API_PORT"
+#        shell: bash
+#
+#      - name: Setup Nix Environment
+#        run: |
+#          echo "Checking for nix installation..."
+#          
+#          # Check if nix binary exists directly
+#          if [ -f /nix/var/nix/profiles/default/bin/nix ]; then
+#            echo "Found nix binary at /nix/var/nix/profiles/default/bin/nix"
+#            export PATH="/nix/var/nix/profiles/default/bin:$PATH"
+#            echo "PATH=$PATH" >> $GITHUB_ENV
+#            nix --version
+#          elif [ -f /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh ]; then
+#            echo "Found nix profile script, sourcing..."
+#            source /nix/var/nix/profiles/default/etc/profile.d/nix-daemon.sh
+#            nix --version
+#          elif command -v nix >/dev/null 2>&1; then
+#            echo "Nix already in PATH"
+#            nix --version
+#          else
+#            echo "Nix not found. Debugging info:"
+#            echo "Contents of /nix/var/nix/profiles/default/:"
+#            ls -la /nix/var/nix/profiles/default/ 2>/dev/null || echo "Directory not found"
+#            echo "Contents of /nix/var/nix/profiles/default/bin/:"
+#            ls -la /nix/var/nix/profiles/default/bin/ 2>/dev/null || echo "Directory not found"
+#            exit 1
+#          fi
+#        shell: bash
+#
+#      - uses: ./.github/actions/lint-check
+#
+#      - uses: ./.github/actions/unit-test
+#
+#      - name: Cleanup EXO_HOME
+#        run: |
+#          echo "Cleaning up EXO_HOME: $EXO_HOME"
+#          rm -rf "$EXO_HOME"
+#        shell: bash
+#        if: always()

.gitignore

@@ -0,0 +1,24 @@
+# gitingest
+digest.txt
+
+# python
+**/__pycache__
+
+# nix
+.direnv/
+
+
+# xcode / macos
+*.xcuserstate
+**/.DS_Store
+
+
+# rust
+target/
+**/*.rs.bk
+*.pdb
+
+# svelte
+dashboard/build/
+dashboard/node_modules/
+dashboard/.svelte-kit/

.idea/.gitignore

@@ -0,0 +1,9 @@
+# Default ignored files
+/shelf/
+/workspace.xml
+# Editor-based HTTP Client requests
+/httpRequests/
+# Datasource local storage ignored files
+/dataSources/
+/dataSources.local.xml
+workspace.xml

.idea/LanguageServersSettings.xml

@@ -0,0 +1,16 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="LanguageServerSettingsState">
+    <state>
+      <map>
+        <entry key="com.insyncwithfoo.pyright">
+          <value>
+            <LanguageServerDefinitionSettings>
+              <option name="errorReportingKind" value="in_log" />
+            </LanguageServerDefinitionSettings>
+          </value>
+        </entry>
+      </map>
+    </state>
+  </component>
+</project>

.idea/exo-v2.iml

@@ -0,0 +1,31 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<module type="EMPTY_MODULE" version="4">
+  <component name="FacetManager">
+    <facet type="Python" name="Python facet">
+      <configuration sdkName="Python 3.13 virtualenv at ~/Desktop/exo/.venv" />
+    </facet>
+  </component>
+  <component name="Go" enabled="true" />
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$">
+      <sourceFolder url="file://$MODULE_DIR$/scripts/src" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/src" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/exo_pyo3_bindings/src" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/exo_pyo3_bindings/tests" isTestSource="true" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/util/src" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/networking/examples" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/networking/src" isTestSource="false" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/networking/tests" isTestSource="true" />
+      <sourceFolder url="file://$MODULE_DIR$/rust/system_custodian/src" isTestSource="false" />
+      <excludeFolder url="file://$MODULE_DIR$/.venv" />
+      <excludeFolder url="file://$MODULE_DIR$/.direnv" />
+      <excludeFolder url="file://$MODULE_DIR$/build" />
+      <excludeFolder url="file://$MODULE_DIR$/dist" />
+      <excludeFolder url="file://$MODULE_DIR$/.go_cache" />
+      <excludeFolder url="file://$MODULE_DIR$/rust/target" />
+    </content>
+    <orderEntry type="jdk" jdkName="Python 3.13 (exo)" jdkType="Python SDK" />
+    <orderEntry type="sourceFolder" forTests="false" />
+    <orderEntry type="library" name="Python 3.13 virtualenv at ~/Desktop/exo/.venv interpreter library" level="application" />
+  </component>
+</module>

.idea/externalDependencies.xml

@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ExternalDependencies">
+    <plugin id="systems.fehn.intellijdirenv" />
+  </component>
+</project>

.idea/inspectionProfiles/Project_Default.xml

@@ -0,0 +1,14 @@
+<component name="InspectionProjectProfileManager">
+  <profile version="1.0">
+    <option name="myName" value="Project Default" />
+    <inspection_tool class="PyCompatibilityInspection" enabled="true" level="WARNING" enabled_by_default="true">
+      <option name="ourVersions">
+        <value>
+          <list size="1">
+            <item index="0" class="java.lang.String" itemvalue="3.14" />
+          </list>
+        </value>
+      </option>
+    </inspection_tool>
+  </profile>
+</component>

.idea/misc.xml

@@ -0,0 +1,10 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="Black">
+    <option name="sdkName" value="Python 3.13 (exo)" />
+  </component>
+  <component name="ProjectRootManager" version="2" project-jdk-name="Python 3.13 (exo)" project-jdk-type="Python SDK" />
+  <component name="PythonCompatibilityInspectionAdvertiser">
+    <option name="version" value="3" />
+  </component>
+</project>

.idea/modules.xml

@@ -0,0 +1,8 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/exo.iml" filepath="$PROJECT_DIR$/.idea/exo.iml" />
+    </modules>
+  </component>
+</project>

.idea/pyright-overrides.xml

@@ -0,0 +1,18 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="com.insyncwithfoo.pyright.configurations.Override">
+    <option name="names">
+      <map>
+        <entry key="configurationFile" value="true" />
+        <entry key="diagnosticMode" value="true" />
+        <entry key="inlayHintsGenericTypes" value="true" />
+        <entry key="prefixTooltipMessages" value="true" />
+        <entry key="runningMode" value="true" />
+        <entry key="smartExecutableResolution" value="true" />
+        <entry key="smartLanguageServerExecutableResolution" value="true" />
+        <entry key="useEditorFontForTooltips" value="true" />
+        <entry key="useTypingExtensions" value="true" />
+      </map>
+    </option>
+  </component>
+</project>

.idea/pyright.xml

@@ -0,0 +1,9 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="com.insyncwithfoo.pyright.configurations.Local">
+    <option name="diagnosticMode" value="WORKSPACE" />
+    <option name="inlayHintsGenericTypes" value="true" />
+    <option name="prefixTooltipMessages" value="true" />
+    <option name="useEditorFontForTooltips" value="true" />
+  </component>
+</project>

.idea/vcs.xml

@@ -0,0 +1,6 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="VcsDirectoryMappings">
+    <mapping directory="" vcs="Git" />
+  </component>
+</project>

.mlx_typings/mlx/core/cuda/__init__.pyi

@@ -0,0 +1,2 @@
+def is_available() -> bool:
+    """Check if the CUDA back-end is available."""

.mlx_typings/mlx/core/distributed/__init__.pyi

@@ -0,0 +1,216 @@
+from typing import Sequence
+
+from mlx.core import Device, Dtype, Stream, array
+
+class Group:
+    """
+    An :class:`mlx.core.distributed.Group` represents a group of independent mlx
+    processes that can communicate.
+    """
+    def rank(self) -> int:
+        """Get the rank of this process"""
+
+    def size(self) -> int:
+        """Get the size of the group"""
+
+    def split(self, color: int, key: int = ...) -> Group:
+        """
+        Split the group to subgroups based on the provided color.
+
+        Processes that use the same color go to the same group. The ``key``
+        argument defines the rank in the new group. The smaller the key the
+        smaller the rank. If the key is negative then the rank in the
+        current group is used.
+
+        Args:
+          color (int): A value to group processes into subgroups.
+          key (int, optional): A key to optionally change the rank ordering
+            of the processes.
+        """
+
+def all_gather(
+    x: array, *, group: Group | None = ..., stream: Stream | Device | None = ...
+) -> array:
+    """
+    Gather arrays from all processes.
+
+    Gather the ``x`` arrays from all processes in the group and concatenate
+    them along the first axis. The arrays should all have the same shape.
+
+    Args:
+      x (array): Input array.
+      group (Group): The group of processes that will participate in the
+        gather. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The concatenation of all ``x`` arrays.
+    """
+
+def all_max(
+    x: array, *, group: Group | None = ..., stream: Stream | Device | None = ...
+) -> array:
+    """
+    All reduce max.
+
+    Find the maximum of the ``x`` arrays from all processes in the group.
+
+    Args:
+      x (array): Input array.
+      group (Group): The group of processes that will participate in the
+        reduction. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The maximum of all ``x`` arrays.
+    """
+
+def all_min(
+    x: array, *, group: Group | None = ..., stream: Stream | Device | None = ...
+) -> array:
+    """
+    All reduce min.
+
+    Find the minimum of the ``x`` arrays from all processes in the group.
+
+    Args:
+      x (array): Input array.
+      group (Group): The group of processes that will participate in the
+        reduction. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The minimum of all ``x`` arrays.
+    """
+
+def all_sum(
+    x: array, *, group: Group | None = ..., stream: Stream | Device | None = ...
+) -> array:
+    """
+    All reduce sum.
+
+    Sum the ``x`` arrays from all processes in the group.
+
+    Args:
+      x (array): Input array.
+      group (Group): The group of processes that will participate in the
+        reduction. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The sum of all ``x`` arrays.
+    """
+
+def init(strict: bool = ..., backend: str = ...) -> Group:
+    """
+    Initialize the communication backend and create the global communication group.
+
+    Example:
+
+      .. code:: python
+
+        import mlx.core as mx
+
+        group = mx.distributed.init(backend="ring")
+
+    Args:
+      strict (bool, optional): If set to False it returns a singleton group
+        in case ``mx.distributed.is_available()`` returns False otherwise
+        it throws a runtime error. Default: ``False``
+      backend (str, optional): Which distributed backend to initialize.
+        Possible values ``mpi``, ``ring``, ``nccl``, ``any``. If set to ``any`` all
+        available backends are tried and the first one that succeeds
+        becomes the global group which will be returned in subsequent
+        calls. Default: ``any``
+
+    Returns:
+      Group: The group representing all the launched processes.
+    """
+
+def is_available() -> bool:
+    """Check if a communication backend is available."""
+
+def recv(
+    shape: Sequence[int],
+    dtype: Dtype,
+    src: int,
+    *,
+    group: Group | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Recv an array with shape ``shape`` and dtype ``dtype`` from process
+    with rank ``src``.
+
+    Args:
+      shape (tuple[int]): The shape of the array we are receiving.
+      dtype (Dtype): The data type of the array we are receiving.
+      src (int): Rank of the source process in the group.
+      group (Group): The group of processes that will participate in the
+        recv. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The array that was received from ``src``.
+    """
+
+def recv_like(
+    x: array,
+    src: int,
+    *,
+    group: Group | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Recv an array with shape and type like ``x`` from process with rank
+    ``src``.
+
+    It is equivalent to calling ``mx.distributed.recv(x.shape, x.dtype, src)``.
+
+    Args:
+      x (array): An array defining the shape and dtype of the array we are
+        receiving.
+      src (int): Rank of the source process in the group.
+      group (Group): The group of processes that will participate in the
+        recv. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: The array that was received from ``src``.
+    """
+
+def send(
+    x: array,
+    dst: int,
+    *,
+    group: Group | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Send an array from the current process to the process that has rank
+    ``dst`` in the group.
+
+    Args:
+      x (array): Input array.
+      dst (int): Rank of the destination process in the group.
+      group (Group): The group of processes that will participate in the
+        sned. If set to ``None`` the global group is used. Default:
+        ``None``.
+      stream (Stream, optional): Stream or device. Defaults to ``None``
+        in which case the default stream of the default device is used.
+
+    Returns:
+      array: An array identical to ``x`` which when evaluated the send is performed.
+    """

.mlx_typings/mlx/core/metal/__init__.pyi

@@ -0,0 +1,38 @@
+def clear_cache() -> None: ...
+def device_info() -> dict[str, str | int]:
+    """
+    Get information about the GPU device and system settings.
+
+    Currently returns:
+
+    * ``architecture``
+    * ``max_buffer_size``
+    * ``max_recommended_working_set_size``
+    * ``memory_size``
+    * ``resource_limit``
+
+    Returns:
+        dict: A dictionary with string keys and string or integer values.
+    """
+
+def get_active_memory() -> int: ...
+def get_cache_memory() -> int: ...
+def get_peak_memory() -> int: ...
+def is_available() -> bool:
+    """Check if the Metal back-end is available."""
+
+def reset_peak_memory() -> None: ...
+def set_cache_limit(limit: int) -> int: ...
+def set_memory_limit(limit: int) -> int: ...
+def set_wired_limit(limit: int) -> int: ...
+def start_capture(path: str) -> None:
+    """
+    Start a Metal capture.
+
+    Args:
+      path (str): The path to save the capture which should have
+        the extension ``.gputrace``.
+    """
+
+def stop_capture() -> None:
+    """Stop a Metal capture."""

.mlx_typings/mlx/core/random/__init__.pyi

@@ -0,0 +1,301 @@
+from typing import Sequence
+
+from mlx.core import Device, Dtype, Stream, array, scalar
+from mlx.core.distributed import state as state
+
+def bernoulli(
+    p: scalar | array = ...,
+    shape: Sequence[int] | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate Bernoulli random values.
+
+    The values are sampled from the bernoulli distribution with parameter
+    ``p``. The parameter ``p`` can be a :obj:`float` or :obj:`array` and
+    must be broadcastable to ``shape``.
+
+    Args:
+        p (float or array, optional): Parameter of the Bernoulli
+          distribution. Default: ``0.5``.
+        shape (list(int), optional): Shape of the output.
+          Default: ``p.shape``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The array of random integers.
+    """
+
+def categorical(
+    logits: array,
+    axis: int = ...,
+    shape: Sequence[int] | None = ...,
+    num_samples: int | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Sample from a categorical distribution.
+
+    The values are sampled from the categorical distribution specified by
+    the unnormalized values in ``logits``. Note, at most one of ``shape``
+    or ``num_samples`` can be specified. If both are ``None``, the output
+    has the same shape as ``logits`` with the ``axis`` dimension removed.
+
+    Args:
+        logits (array): The *unnormalized* categorical distribution(s).
+        axis (int, optional): The axis which specifies the distribution.
+           Default: ``-1``.
+        shape (list(int), optional): The shape of the output. This must
+           be broadcast compatible with ``logits.shape`` with the ``axis``
+           dimension removed. Default: ``None``
+        num_samples (int, optional): The number of samples to draw from each
+          of the categorical distributions in ``logits``. The output will have
+          ``num_samples`` in the last dimension. Default: ``None``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The ``shape``-sized output array with type ``uint32``.
+    """
+
+def gumbel(
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    key: Stream | Device | None = ...,
+    stream: array | None = ...,
+) -> array:
+    """
+    Sample from the standard Gumbel distribution.
+
+    The values are sampled from a standard Gumbel distribution
+    which CDF ``exp(-exp(-x))``.
+
+    Args:
+        shape (list(int)): The shape of the output.
+        dtype (Dtype, optional): The data type of the output.
+          Default: ``float32``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array:
+          The :class:`array` with shape ``shape`` and distributed according
+          to the Gumbel distribution.
+    """
+
+def key(seed: int) -> array:
+    """
+    Get a PRNG key from a seed.
+
+    Args:
+        seed (int): Seed for the PRNG.
+
+    Returns:
+        array: The PRNG key array.
+    """
+
+def laplace(
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    loc: float = ...,
+    scale: float = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Sample numbers from a Laplace distribution.
+
+    Args:
+        shape (list(int), optional): Shape of the output. Default: ``()``.
+        dtype (Dtype, optional): Type of the output. Default: ``float32``.
+        loc (float, optional): Mean of the distribution. Default: ``0.0``.
+        scale (float, optional): The scale "b" of the Laplace distribution.
+          Default:``1.0``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The output array of random values.
+    """
+
+def multivariate_normal(
+    mean: array,
+    cov: array,
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate jointly-normal random samples given a mean and covariance.
+
+    The matrix ``cov`` must be positive semi-definite. The behavior is
+    undefined if it is not.  The only supported ``dtype`` is ``float32``.
+
+    Args:
+        mean (array): array of shape ``(..., n)``, the mean of the
+          distribution.
+        cov (array): array  of shape ``(..., n, n)``, the covariance
+          matrix of the distribution. The batch shape ``...`` must be
+          broadcast-compatible with that of ``mean``.
+        shape (list(int), optional): The output shape must be
+          broadcast-compatible with ``mean.shape[:-1]`` and ``cov.shape[:-2]``.
+          If empty, the result shape is determined by broadcasting the batch
+          shapes of ``mean`` and ``cov``. Default: ``[]``.
+        dtype (Dtype, optional): The output type. Default: ``float32``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The output array of random values.
+    """
+
+def normal(
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    loc: scalar | array | None = ...,
+    scale: scalar | array | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    r"""
+    Generate normally distributed random numbers.
+
+    If ``loc`` and ``scale`` are not provided the "standard" normal
+    distribution is used. That means $x \sim \mathcal{N}(0, 1)$ for
+    real numbers and $\text{Re}(x),\text{Im}(x) \sim \mathcal{N}(0,
+    \frac{1}{2})$ for complex numbers.
+
+    Args:
+        shape (list(int), optional): Shape of the output. Default: ``()``.
+        dtype (Dtype, optional): Type of the output. Default: ``float32``.
+        loc (scalar or array, optional): Mean of the distribution.
+          Default: ``None``.
+        scale (scalar or array, optional): Standard deviation of the
+          distribution. Default: ``None``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The output array of random values.
+    """
+
+def permutation(
+    x: int | array,
+    axis: int = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate a random permutation or permute the entries of an array.
+
+    Args:
+        x (int or array, optional): If an integer is provided a random
+          permtuation of ``mx.arange(x)`` is returned. Otherwise the entries
+          of ``x`` along the given axis are randomly permuted.
+        axis (int, optional): The axis to permute along. Default: ``0``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array:
+          The generated random permutation or randomly permuted input array.
+    """
+
+def randint(
+    low: scalar | array,
+    high: scalar | array,
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate random integers from the given interval.
+
+    The values are sampled with equal probability from the integers in
+    half-open interval ``[low, high)``. The lower and upper bound can be
+    scalars or arrays and must be broadcastable to ``shape``.
+
+    Args:
+        low (scalar or array): Lower bound of the interval.
+        high (scalar or array): Upper bound of the interval.
+        shape (list(int), optional): Shape of the output. Default: ``()``.
+        dtype (Dtype, optional): Type of the output. Default: ``int32``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The array of random integers.
+    """
+
+def seed(seed: int) -> None:
+    """
+    Seed the global PRNG.
+
+    Args:
+        seed (int): Seed for the global PRNG.
+    """
+
+def split(key: array, num: int = ..., stream: Stream | Device | None = ...) -> array:
+    """
+    Split a PRNG key into sub keys.
+
+    Args:
+        key (array): Input key to split.
+        num (int, optional): Number of sub keys. Default: ``2``.
+
+    Returns:
+        array: The array of sub keys with ``num`` as its first dimension.
+    """
+
+def truncated_normal(
+    lower: scalar | array,
+    upper: scalar | array,
+    shape: Sequence[int] | None = ...,
+    dtype: Dtype | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate values from a truncated normal distribution.
+
+    The values are sampled from the truncated normal distribution
+    on the domain ``(lower, upper)``. The bounds ``lower`` and ``upper``
+    can be scalars or arrays and must be broadcastable to ``shape``.
+
+    Args:
+        lower (scalar or array): Lower bound of the domain.
+        upper (scalar or array): Upper bound of the domain.
+        shape (list(int), optional): The shape of the output.
+          Default:``()``.
+        dtype (Dtype, optional): The data type of the output.
+          Default: ``float32``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The output array of random values.
+    """
+
+def uniform(
+    low: scalar | array = ...,
+    high: scalar | array = ...,
+    shape: Sequence[int] = ...,
+    dtype: Dtype | None = ...,
+    key: array | None = ...,
+    stream: Stream | Device | None = ...,
+) -> array:
+    """
+    Generate uniformly distributed random numbers.
+
+    The values are sampled uniformly in the half-open interval ``[low, high)``.
+    The lower and upper bound can be scalars or arrays and must be
+    broadcastable to ``shape``.
+
+    Args:
+        low (scalar or array, optional): Lower bound of the distribution.
+          Default: ``0``.
+        high (scalar or array, optional): Upper bound of the distribution.
+          Default: ``1``.
+        shape (list(int), optional): Shape of the output. Default:``()``.
+        dtype (Dtype, optional): Type of the output. Default: ``float32``.
+        key (array, optional): A PRNG key. Default: ``None``.
+
+    Returns:
+        array: The output array random values.
+    """

.mlx_typings/mlx/nn/__init__.pyi

@@ -0,0 +1,9 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from layers import *
+from utils import *
+
+from . import init as init
+from . import losses as losses

.mlx_typings/mlx/nn/init.pyi

@@ -0,0 +1,284 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Callable, Literal
+
+import mlx.core as mx
+
+def constant(value: float, dtype: mx.Dtype = ...) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns an array filled with ``value``.
+
+    Args:
+        value (float): The value to fill the array with.
+        dtype (Dtype, optional): The data type of the array. Default:
+          ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns an array with the
+        same shape as the input, filled with ``value``.
+
+    Example:
+
+        >>> init_fn = nn.init.constant(0.5)
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[0.5, 0.5],
+               [0.5, 0.5]], dtype=float32)
+    """
+
+def normal(
+    mean: float = ..., std: float = ..., dtype: mx.Dtype = ...
+) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns samples from a normal distribution.
+
+    Args:
+        mean (float, optional): Mean of the normal distribution. Default:
+          ``0.0``.
+        std (float, optional): Standard deviation of the normal distribution.
+          Default: ``1.0``.
+        dtype (Dtype, optional): The data type of the array. Default:
+          ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns an array with the
+        same shape as the input, filled with samples from a normal distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.normal()
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[-0.982273, -0.534422],
+               [0.380709, 0.0645099]], dtype=float32)
+    """
+
+def uniform(
+    low: float = ..., high: float = ..., dtype: mx.Dtype = ...
+) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns samples from a uniform distribution.
+
+    Args:
+        low (float, optional): The lower bound of the uniform distribution.
+          Default: ``0.0``.
+        high (float, optional): The upper bound of the uniform distribution.
+          Default: ``1.0``
+        dtype (Dtype, optional): The data type of the array. Default: ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns an array
+        with the same shape as the input, filled with samples from a uniform
+        distribution
+
+    Example:
+
+        >>> init_fn = nn.init.uniform(low=0, high=1)
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[0.883935, 0.863726],
+               [0.617261, 0.417497]], dtype=float32)
+    """
+
+def identity(dtype: mx.Dtype = ...) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns an identity matrix.
+
+    Args:
+        dtype (Dtype, optional): The data type of the array. Defaults:
+          ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns an identity
+        matrix with the same shape as the input.
+
+    Example:
+
+        >>> init_fn = nn.init.identity()
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[1, 0],
+               [0, 1]], dtype=float32)
+    """
+
+def glorot_normal(dtype: mx.Dtype = ...) -> Callable[[mx.array, float], mx.array]:
+    r"""A Glorot normal initializer.
+
+    This initializer samples from a normal distribution with a standard
+    deviation computed from the number of input (``fan_in``) and output
+    (``fan_out``) units according to:
+
+    .. math::
+        \sigma = \gamma \sqrt{\frac{2.0}{\text{fan\_in} + \text{fan\_out}}}
+
+    For more details see the original reference: `Understanding the difficulty
+    of training deep feedforward neural networks
+    <https://proceedings.mlr.press/v9/glorot10a.html>`_
+
+    Args:
+        dtype (Dtype, optional): The data type of the array. Default: ``float32``.
+
+    Returns:
+        Callable[[array, float], array]: An initializer that returns an array
+        with the same shape as the input, filled with samples from the Glorot
+        normal distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.glorot_normal()
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[0.191107, 1.61278],
+               [-0.150594, -0.363207]], dtype=float32)
+        >>> init_fn(mx.zeros((2, 2)), gain=4.0)
+        array([[1.89613, -4.53947],
+               [4.48095, 0.995016]], dtype=float32)
+    """
+
+def glorot_uniform(dtype: mx.Dtype = ...) -> Callable[[mx.array, float], mx.array]:
+    r"""A Glorot uniform initializer.
+
+    This initializer samples from a uniform distribution with a range
+    computed from the number of input (``fan_in``) and output (``fan_out``)
+    units according to:
+
+    .. math::
+        \sigma = \gamma \sqrt{\frac{6.0}{\text{fan\_in} + \text{fan\_out}}}
+
+    For more details see the original reference: `Understanding the difficulty
+    of training deep feedforward neural networks
+    <https://proceedings.mlr.press/v9/glorot10a.html>`_
+
+    Args:
+        dtype (Dtype, optional): The data type of the array. Default: ``float32``.
+
+    Returns:
+        Callable[[array, float], array]: An initializer that returns an array
+        with the same shape as the input, filled with samples from the Glorot
+        uniform distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.glorot_uniform()
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[0.223404, -0.890597],
+               [-0.379159, -0.776856]], dtype=float32)
+        >>> init_fn(mx.zeros((2, 2)), gain=4.0)
+        array([[-1.90041, 3.02264],
+               [-0.912766, 4.12451]], dtype=float32)
+    """
+
+def he_normal(
+    dtype: mx.Dtype = ...,
+) -> Callable[[mx.array, Literal["fan_in", "fan_out"], float], mx.array]:
+    r"""Build a He normal initializer.
+
+    This initializer samples from a normal distribution with a standard
+    deviation computed from the number of input (``fan_in``) or output
+    (``fan_out``) units according to:
+
+    .. math::
+        \sigma = \gamma \frac{1}{\sqrt{\text{fan}}}
+
+    where :math:`\text{fan}` is either the number of input units when the
+    ``mode`` is ``"fan_in"`` or output units when the ``mode`` is
+    ``"fan_out"``.
+
+    For more details see the original reference: `Delving Deep into Rectifiers:
+    Surpassing Human-Level Performance on ImageNet Classification
+    <https://arxiv.org/abs/1502.01852>`_
+
+    Args:
+        dtype (Dtype, optional): The data type of the array. Defaults to mx.float32.
+
+    Returns:
+        Callable[[array, str, float], array]: An initializer that returns an
+        array with the same shape as the input, filled with samples from the He
+        normal distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.he_normal()
+        >>> init_fn(mx.zeros((2, 2)))  # uses fan_in
+        array([[-1.25211, 0.458835],
+               [-0.177208, -0.0137595]], dtype=float32)
+        >>> init_fn(mx.zeros((2, 2)), mode="fan_out", gain=5)
+        array([[5.6967, 4.02765],
+               [-4.15268, -2.75787]], dtype=float32)
+    """
+
+def he_uniform(
+    dtype: mx.Dtype = ...,
+) -> Callable[[mx.array, Literal["fan_in", "fan_out"], float], mx.array]:
+    r"""A He uniform (Kaiming uniform) initializer.
+
+    This initializer samples from a uniform distribution with a range
+    computed from the number of input (``fan_in``) or output (``fan_out``)
+    units according to:
+
+    .. math::
+
+        \sigma = \gamma \sqrt{\frac{3.0}{\text{fan}}}
+
+    where :math:`\text{fan}` is either the number of input units when the
+    ``mode`` is ``"fan_in"`` or output units when the ``mode`` is
+    ``"fan_out"``.
+
+    For more details see the original reference: `Delving Deep into Rectifiers:
+    Surpassing Human-Level Performance on ImageNet Classification
+    <https://arxiv.org/abs/1502.01852>`_
+
+
+    Args:
+        dtype (Dtype, optional): The data type of the array. Default: ``float32``.
+
+    Returns:
+        Callable[[array, str, float], array]: An initializer that returns an
+        array with the same shape as the input, filled with samples from  the
+        He uniform distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.he_uniform()
+        >>> init_fn(mx.zeros((2, 2)))  # uses fan_in
+        array([[0.0300242, -0.0184009],
+               [0.793615, 0.666329]], dtype=float32)
+        >>> init_fn(mx.zeros((2, 2)), mode="fan_out", gain=5)
+        array([[-1.64331, -2.16506],
+               [1.08619, 5.79854]], dtype=float32)
+    """
+
+def sparse(
+    sparsity: float, mean: float = ..., std: float = ..., dtype: mx.Dtype = ...
+) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns a sparse matrix.
+
+    Args:
+        sparsity (float): The fraction of elements in each column to be set to
+        zero.
+        mean (float, optional): Mean of the normal distribution. Default:
+          ``0.0``.
+        std (float, optional): Standard deviation of the normal distribution.
+          Default: ``1.0``.
+        dtype (Dtype, optional): The data type of the array. Default:
+          ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns an array with the
+        same shape as the input, filled with samples from a normal distribution.
+
+    Example:
+
+        >>> init_fn = nn.init.sparse(sparsity=0.5)
+        >>> init_fn(mx.zeros((2, 2)))
+        array([[-1.91187, -0.117483],
+       [0, 0]], dtype=float32)
+    """
+
+def orthogonal(
+    gain: float = ..., dtype: mx.Dtype = ...
+) -> Callable[[mx.array], mx.array]:
+    r"""An initializer that returns an orthogonal matrix.
+
+    Args:
+        gain (float, optional): Scaling factor for the orthogonal matrix.
+            Default: ``1.0``.
+        dtype (Dtype, optional): Data type of the array. Default: ``float32``.
+
+    Returns:
+        Callable[[array], array]: An initializer that returns
+        an orthogonal matrix with the same shape as the input.
+    """

.mlx_typings/mlx/nn/layers/__init__.pyi

@@ -0,0 +1,20 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from activations import *
+from base import *
+from containers import *
+from convolution import *
+from convolution_transpose import *
+from distributed import *
+from dropout import *
+from embedding import *
+from linear import *
+from normalization import *
+from pooling import *
+from positional_encoding import *
+from quantized import *
+from recurrent import *
+from transformer import *
+from upsample import *

.mlx_typings/mlx/nn/layers/activations.pyi

@@ -0,0 +1,523 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from functools import partial
+from typing import Any
+
+import mlx.core as mx
+from base import Module
+
+@partial(mx.compile, shapeless=True)
+def sigmoid(x: mx.array) -> mx.array:
+    r"""Applies the sigmoid function.
+
+    .. math::
+        \text{Sigmoid}(x) = \sigma(x) = \frac{1}{1 + \exp(-x)}
+    """
+
+@partial(mx.compile, shapeless=True)
+def relu(x: mx.array) -> mx.array:
+    r"""Applies the Rectified Linear Unit.
+
+    Simply ``mx.maximum(x, 0)``.
+    """
+
+@partial(mx.compile, shapeless=True)
+def relu2(x: mx.array) -> mx.array:
+    r"""Applies the ReLU² activation function.
+
+    Applies :math:`\max(0, x)^2` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def relu6(x: mx.array) -> mx.array:
+    r"""Applies the Rectified Linear Unit 6.
+
+    Applies :math:`\min(\max(x, 0), 6)` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def leaky_relu(x: mx.array, negative_slope=...) -> mx.array:
+    r"""Applies the Leaky Rectified Linear Unit.
+
+    Simply ``mx.maximum(negative_slope * x, x)``.
+    """
+
+@partial(mx.compile, shapeless=True)
+def log_softmax(x: mx.array, axis=...):
+    r"""Applies the Log Softmax function.
+
+    Applies :math:`x + \log \sum_i e^{x_i}` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def elu(x: mx.array, alpha=...) -> mx.array:
+    r"""Applies the Exponential Linear Unit.
+
+    Simply ``mx.where(x > 0, x, alpha * (mx.exp(x) - 1))``.
+    """
+
+@partial(mx.compile, shapeless=True)
+def softmax(x: mx.array, axis=...) -> mx.array:
+    r"""Applies the Softmax function.
+
+    Applies :math:`\frac{e^{x_i}}{\sum_j e^{x_j}}` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def softplus(x: mx.array) -> mx.array:
+    r"""Applies the Softplus function.
+
+    Applies :math:`\log(1 + \exp(x))` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def softsign(x: mx.array) -> mx.array:
+    r"""Applies the Softsign function.
+
+    Applies :math:`\frac{x}{1 + |x|}` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def softshrink(x: mx.array, lambd: float = ...) -> mx.array:
+    r"""Applies the Softshrink activation function.
+
+    .. math::
+        \text{softshrink}(x) = \begin{cases}
+        x - \lambda & \text{if } x > \lambda \\
+        x + \lambda & \text{if } x < -\lambda \\
+        0 & \text{otherwise}
+        \end{cases}
+    """
+
+@partial(mx.compile, shapeless=True)
+def celu(x: mx.array, alpha=...) -> mx.array:
+    r"""Applies the Continuously Differentiable Exponential Linear Unit.
+
+    Applies :math:`\max(0, x) + \min(0, \alpha * (\exp(x / \alpha) - 1))`
+    element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def silu(x: mx.array) -> mx.array:
+    r"""Applies the Sigmoid Linear Unit. Also known as Swish.
+
+    Applies :math:`x \sigma(x)` element wise, where :math:`\sigma(\cdot)` is
+    the logistic sigmoid.
+    """
+
+@partial(mx.compile, shapeless=True)
+def log_sigmoid(x: mx.array) -> mx.array:
+    r"""Applies the Log Sigmoid function.
+
+    Applies :math:`\log(\sigma(x)) = -\log(1 + e^{-x})` element wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def gelu(x: mx.array) -> mx.array:
+    r"""Applies the Gaussian Error Linear Units function.
+
+    .. math::
+        \textrm{GELU}(x) = x * \Phi(x)
+
+    where :math:`\Phi(x)` is the Gaussian CDF.
+
+    See also :func:`gelu_approx` and :func:`gelu_fast_approx` for faster
+    approximations.
+    """
+
+@partial(mx.compile, shapeless=True)
+def gelu_approx(x: mx.array) -> mx.array:
+    r"""An approximation to Gaussian Error Linear Unit.
+
+    See :func:`gelu` for the exact computation.
+
+    This function approximates ``gelu`` with a maximum absolute error :math:`<
+    0.0005` in the range :math:`[-6, 6]` using the following
+
+    .. math::
+
+        x = 0.5 * x * \left(1 + \text{Tanh}\left((\sqrt{2 / \pi} * \left(x + 0.044715 * x^3\right)\right)\right)
+
+    """
+
+@partial(mx.compile, shapeless=True)
+def gelu_fast_approx(x: mx.array) -> mx.array:
+    r"""A fast approximation to Gaussian Error Linear Unit.
+
+    See :func:`gelu` for the exact computation.
+
+    This function approximates ``gelu`` with a maximum absolute error :math:`<
+    0.015` in the range :math:`[-6, 6]` using the following
+
+    .. math::
+
+        x = x \sigma\left(1.702 x\right)
+
+    where :math:`\sigma(\cdot)` is the logistic sigmoid.
+
+    References:
+    - https://github.com/hendrycks/GELUs
+    - https://arxiv.org/abs/1606.08415
+    """
+
+def glu(x: mx.array, axis: int = ...) -> mx.array:
+    r"""Applies the gated linear unit function.
+
+    This function splits the ``axis`` dimension of the input into two halves
+    (:math:`a` and :math:`b`) and applies :math:`a * \sigma(b)`.
+
+    .. math::
+        \textrm{GLU}(x) = a * \sigma(b)
+
+    Args:
+        axis (int): The dimension to split along. Default: ``-1``
+    """
+
+@partial(mx.compile, shapeless=True)
+def step(x: mx.array, threshold: float = ...) -> mx.array:
+    r"""Applies the Step Activation Function.
+
+    This function implements a binary step activation, where the output is set
+    to 1 if the input is greater than a specified threshold, and 0 otherwise.
+
+    .. math::
+        \text{step}(x) = \begin{cases}
+        0 & \text{if } x < \text{threshold} \\
+        1 & \text{if } x \geq \text{threshold}
+        \end{cases}
+
+    Args:
+        threshold: The value to threshold at.
+    """
+
+@partial(mx.compile, shapeless=True)
+def selu(x: mx.array) -> mx.array:
+    r"""Applies the Scaled Exponential Linear Unit.
+
+    .. math::
+        \text{selu}(x) = \begin{cases}
+        \lambda x & \text{if } x > 0 \\
+        \lambda \alpha (\exp(x) - 1) & \text{if } x \leq 0
+        \end{cases}
+
+    where :math:`\lambda = 1.0507` and :math:`\alpha = 1.67326`.
+
+    See also :func:`elu`.
+    """
+
+@partial(mx.compile, shapeless=True)
+def prelu(x: mx.array, alpha: mx.array) -> mx.array:
+    r"""Applies the element-wise parametric ReLU.
+
+    .. math::
+        \text{PReLU}(x) = \max(0,x) + a * \min(0,x)
+
+    where :math:`a` is an array.
+    """
+
+@partial(mx.compile, shapeless=True)
+def mish(x: mx.array) -> mx.array:
+    r"""Applies the Mish function, element-wise.
+
+    Mish: A Self Regularized Non-Monotonic Neural Activation Function.
+
+    Reference: https://arxiv.org/abs/1908.08681
+
+    .. math::
+        \text{Mish}(x) = x * \text{Tanh}(\text{Softplus}(x))
+
+    """
+
+@partial(mx.compile, shapeless=True)
+def hardswish(x: mx.array) -> mx.array:
+    r"""Applies the hardswish function, element-wise.
+
+    .. math::
+        \text{Hardswish}(x) = x * \min(\max(x + 3, 0), 6) / 6
+    """
+
+@partial(mx.compile, shapeless=True)
+def hard_tanh(x: mx.array, min_val=..., max_val=...) -> mx.array:
+    r"""Applies the HardTanh function.
+
+    Applies :math:`\max(\min(x, \text{max\_val}), \text{min\_val})` element-wise.
+    """
+
+@partial(mx.compile, shapeless=True)
+def hard_shrink(x: mx.array, lambd=...) -> mx.array:
+    r"""Applies the HardShrink activation function.
+
+    .. math::
+        \text{hardshrink}(x) = \begin{cases}
+        x & \text{if } x > \lambda \\
+        x & \text{if } x < -\lambda \\
+        0 & \text{otherwise}
+        \end{cases}
+    """
+
+@partial(mx.compile, shapeless=True)
+def softmin(x: mx.array, axis=...) -> mx.array:
+    r"""Applies the Softmin function.
+
+    Applies :math:`\frac{e^{-x_i}}{\sum_j e^{-x_j}}` element-wise.
+    """
+
+def tanh(x: mx.array) -> mx.array:
+    """Applies the hyperbolic tangent function.
+
+    Simply ``mx.tanh(x)``.
+    """
+
+class GLU(Module):
+    r"""Applies the gated linear unit function.
+
+    This function splits the ``axis`` dimension of the input into two halves
+    (:math:`a` and :math:`b`) and applies :math:`a * \sigma(b)`.
+
+    .. math::
+        \textrm{GLU}(x) = a * \sigma(b)
+
+    Args:
+        axis (int): The dimension to split along. Default: ``-1``
+    """
+    def __init__(self, axis: int = ...) -> None: ...
+    def __call__(self, x) -> Any: ...
+
+@_make_activation_module(sigmoid)
+class Sigmoid(Module):
+    r"""Applies the sigmoid function, element-wise.
+
+    .. math::
+        \text{Sigmoid}(x) = \sigma(x) = \frac{1}{1 + \exp(-x)}
+    """
+
+@_make_activation_module(mish)
+class Mish(Module):
+    r"""Applies the Mish function, element-wise.
+
+    Reference: https://arxiv.org/abs/1908.08681
+
+    .. math::
+        \text{Mish}(x) = x * \text{Tanh}(\text{Softplus}(x))
+
+    """
+
+@_make_activation_module(relu)
+class ReLU(Module):
+    r"""Applies the Rectified Linear Unit.
+        Simply ``mx.maximum(x, 0)``.
+
+    See :func:`relu` for the functional equivalent.
+    """
+
+@_make_activation_module(relu2)
+class ReLU2(Module):
+    r"""Applies the ReLU² activation function.
+
+    See :func:`relu2` for the functional equivalent.
+    """
+
+@_make_activation_module(relu6)
+class ReLU6(Module):
+    r"""Applies the Rectified Linear Unit 6.
+
+    See :func:`relu6` for the functional equivalent.
+    """
+
+class LeakyReLU(Module):
+    r"""Applies the Leaky Rectified Linear Unit.
+
+    Simply ``mx.maximum(negative_slope * x, x)``.
+
+    Args:
+        negative_slope: Controls the angle of the negative slope. Default: ``1e-2``
+    """
+    def __init__(self, negative_slope=...) -> None: ...
+    def __call__(self, x): ...
+
+class ELU(Module):
+    r"""Applies the Exponential Linear Unit.
+        Simply ``mx.where(x > 0, x, alpha * (mx.exp(x) - 1))``.
+
+    See :func:`elu` for the functional equivalent.
+
+    Args:
+        alpha: the :math:`\alpha` value for the ELU formulation. Default: ``1.0``
+    """
+    def __init__(self, alpha=...) -> None: ...
+    def __call__(self, x): ...
+
+@_make_activation_module(softmax)
+class Softmax(Module):
+    r"""Applies the Softmax function.
+
+    See :func:`softmax` for the functional equivalent.
+    """
+
+@_make_activation_module(softplus)
+class Softplus(Module):
+    r"""Applies the Softplus function.
+
+    See :func:`softplus` for the functional equivalent.
+    """
+
+@_make_activation_module(softsign)
+class Softsign(Module):
+    r"""Applies the Softsign function.
+
+    See :func:`softsign` for the functional equivalent.
+    """
+
+class Softshrink(Module):
+    r"""Applies the Softshrink function.
+
+    See :func:`softshrink` for the functional equivalent.
+
+    Args:
+        lambd: the :math:`\lambda` value for Softshrink. Default: ``0.5``
+    """
+    def __init__(self, lambd=...) -> None: ...
+    def __call__(self, x): ...
+
+class CELU(Module):
+    r"""Applies the Continuously Differentiable Exponential Linear Unit.
+        Applies :math:`\max(0, x) + \min(0, \alpha * (\exp(x / \alpha) - 1))`
+        element wise.
+
+    See :func:`celu` for the functional equivalent.
+
+    Args:
+        alpha: the :math:`\alpha` value for the CELU formulation. Default: ``1.0``
+    """
+    def __init__(self, alpha=...) -> None: ...
+    def __call__(self, x): ...
+
+@_make_activation_module(silu)
+class SiLU(Module):
+    r"""Applies the Sigmoid Linear Unit. Also known as Swish.
+
+    See :func:`silu` for the functional equivalent.
+    """
+
+@_make_activation_module(log_softmax)
+class LogSoftmax(Module):
+    r"""Applies the Log Softmax function.
+
+    See :func:`log_softmax` for the functional equivalent.
+    """
+
+@_make_activation_module(log_sigmoid)
+class LogSigmoid(Module):
+    r"""Applies the Log Sigmoid function.
+
+    See :func:`log_sigmoid` for the functional equivalent.
+    """
+
+class PReLU(Module):
+    r"""Applies the element-wise parametric ReLU.
+        Applies :math:`\max(0, x) + a * \min(0, x)` element wise, where :math:`a`
+        is an array.
+
+    See :func:`prelu` for the functional equivalent.
+
+    Args:
+        num_parameters: number of :math:`a` to learn. Default: ``1``
+        init: the initial value of :math:`a`. Default: ``0.25``
+    """
+    def __init__(self, num_parameters=..., init=...) -> None: ...
+    def __call__(self, x: mx.array): ...
+
+class GELU(Module):
+    r"""Applies the Gaussian Error Linear Units.
+
+    .. math::
+        \textrm{GELU}(x) = x * \Phi(x)
+
+    where :math:`\Phi(x)` is the Gaussian CDF.
+
+    However, if ``approx`` is set to 'precise' or 'fast' it applies
+
+    .. math::
+        \textrm{GELUApprox}(x) &= 0.5 * x * \left(1 + \text{Tanh}\left((\sqrt{2 / \pi} * \left(x + 0.044715 * x^3\right)\right)\right) \\
+        \textrm{GELUFast}(x) &= x * \sigma\left(1.702 * x\right)
+
+    respectively.
+
+    .. note::
+       For compatibility with the PyTorch API, 'tanh' can be used as an alias
+       for 'precise'.
+
+    See :func:`gelu`, :func:`gelu_approx` and :func:`gelu_fast_approx` for the
+    functional equivalents and information regarding error bounds.
+
+
+    Args:
+        approx ('none' | 'precise' | 'fast'): Which approximation to gelu to use if any.
+    """
+    def __init__(self, approx=...) -> None: ...
+    def __call__(self, x): ...
+
+@_make_activation_module(tanh)
+class Tanh(Module):
+    r"""Applies the hyperbolic tangent function.
+
+    See :func:`tanh` for the functional equivalent.
+    """
+
+@_make_activation_module(hardswish)
+class Hardswish(Module):
+    r"""Applies the hardswish function, element-wise.
+
+    See :func:`hardswish` for the functional equivalent.
+    """
+
+class Step(Module):
+    r"""Applies the Step Activation Function.
+
+    This function implements a binary step activation, where the output is set
+    to 1 if the input is greater than a specified threshold, and 0 otherwise.
+
+    .. math::
+        \text{step}(x) = \begin{cases}
+        0 & \text{if } x < \text{threshold} \\
+        1 & \text{if } x \geq \text{threshold}
+        \end{cases}
+
+    Args:
+        threshold: The value to threshold at.
+    """
+    def __init__(self, threshold: float = ...) -> None: ...
+    def __call__(self, x: mx.array): ...
+
+@_make_activation_module(selu)
+class SELU(Module):
+    r"""Applies the Scaled Exponential Linear Unit.
+
+    See :func:`selu` for the functional equivalent.
+    """
+
+@_make_activation_module(hard_tanh)
+class HardTanh(Module):
+    r"""Applies the HardTanh function.
+
+    See :func:`hard_tanh` for the functional equivalent.
+    """
+
+@_make_activation_module(hard_shrink)
+class HardShrink(Module):
+    r"""Applies the HardShrink function.
+
+    See :func:`hard_shrink` for the functional equivalent.
+
+    Args:
+        lambd: the :math:`\lambda` value for Hardshrink. Default: ``0.5``
+    """
+
+@_make_activation_module(softmin)
+class Softmin(Module):
+    r"""Applies the Softmin function.
+
+    See :func:`softmin` for the functional equivalent.
+    """

.mlx_typings/mlx/nn/layers/base.pyi

@@ -0,0 +1,393 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any, Callable, List, Optional, Tuple, Union
+
+import mlx.core as mx
+
+class Module(dict):
+    """Base class for building neural networks with MLX.
+
+    All the layers provided in :mod:`layers` subclass this class and
+    your models should do the same.
+
+    A ``Module`` can contain other ``Module`` instances or :class:`mlx.core.array`
+    instances in arbitrary nesting of python lists or dicts. The ``Module``
+    then allows recursively extracting all the :class:`mlx.core.array` instances
+    using :meth:`Module.parameters`.
+
+    In addition, the ``Module`` has the concept of trainable and non trainable
+    parameters (called "frozen"). When using :func:`value_and_grad`
+    the gradients are returned only with respect to the trainable parameters.
+    All arrays in a module are trainable unless they are added in the "frozen"
+    set by calling :meth:`freeze`.
+
+    .. code-block:: python
+
+        import mlx.core as mx
+        import mlx.nn as nn
+
+        class MyMLP(nn.Module):
+            def __init__(self, in_dims: int, out_dims: int, hidden_dims: int = 16):
+                super().__init__()
+
+                self.in_proj = nn.Linear(in_dims, hidden_dims)
+                self.out_proj = nn.Linear(hidden_dims, out_dims)
+
+            def __call__(self, x):
+                x = self.in_proj(x)
+                x = mx.maximum(x, 0)
+                return self.out_proj(x)
+
+        model = MyMLP(2, 1)
+
+        # All the model parameters are created but since MLX is lazy by
+        # default, they are not evaluated yet. Calling `mx.eval` actually
+        # allocates memory and initializes the parameters.
+        mx.eval(model.parameters())
+
+        # Setting a parameter to a new value is as simply as accessing that
+        # parameter and assigning a new array to it.
+        model.in_proj.weight = model.in_proj.weight * 2
+        mx.eval(model.parameters())
+    """
+
+    __call__: Callable
+    def __init__(self) -> None:
+        """Should be called by the subclasses of ``Module``."""
+
+    @property
+    def training(self):  # -> bool:
+        """Boolean indicating if the model is in training mode."""
+
+    @property
+    def state(self):  # -> Self:
+        """The module's state dictionary
+
+        The module's state dictionary contains any attribute set on the
+        module including parameters in :meth:`Module.parameters`
+
+        Unlike :meth:`Module.parameters`, the :attr:`Module.state` property is
+        a reference to the module's state. Updates to it will be reflected in
+        the original module.
+        """
+
+    def __repr__(self):  # -> str:
+        ...
+    def __getattr__(self, key: str):  # -> None:
+        ...
+    def __setattr__(self, key: str, val: Any):  # -> None:
+        ...
+    def __delattr__(self, name):  # -> None:
+        ...
+    def load_weights(
+        self,
+        file_or_weights: Union[str, List[Tuple[str, mx.array]]],
+        strict: bool = ...,
+    ) -> Module:
+        """
+        Update the model's weights from a ``.npz``, a ``.safetensors`` file, or a list.
+
+        Args:
+            file_or_weights (str or list(tuple(str, mx.array))): The path to
+                the weights ``.npz`` file (``.npz`` or ``.safetensors``) or a list
+                of pairs of parameter names and arrays.
+            strict (bool, optional): If ``True`` then checks that the provided
+              weights exactly match the parameters of the model. Otherwise,
+              only the weights actually contained in the model are loaded and
+              shapes are not checked. Default: ``True``.
+
+        Returns:
+            The module instance after updating the weights.
+
+        Example:
+
+            .. code-block:: python
+
+                import mlx.core as mx
+                import mlx.nn as nn
+                model = nn.Linear(10, 10)
+
+                # Load from file
+                model.load_weights("weights.npz")
+
+                # Load from .safetensors file
+                model.load_weights("weights.safetensors")
+
+                # Load from list
+                weights = [
+                    ("weight", mx.random.uniform(shape=(10, 10))),
+                    ("bias",  mx.zeros((10,))),
+                ]
+                model.load_weights(weights)
+
+                # Missing weight
+                weights = [
+                    ("weight", mx.random.uniform(shape=(10, 10))),
+                ]
+
+                # Raises a ValueError exception
+                model.load_weights(weights)
+
+                # Ok, only updates the weight but not the bias
+                model.load_weights(weights, strict=False)
+        """
+
+    def save_weights(self, file: str):  # -> None:
+        """
+        Save the model's weights to a file. The saving method is determined by the file extension:
+        - ``.npz`` will use :func:`mx.savez`
+        - ``.safetensors`` will use :func:`mx.save_safetensors`
+        """
+
+    @staticmethod
+    def is_module(value):  # -> bool:
+        ...
+    @staticmethod
+    def valid_child_filter(module, key, value):  # -> bool:
+        ...
+    @staticmethod
+    def valid_parameter_filter(module, key, value):  # -> bool:
+        ...
+    @staticmethod
+    def trainable_parameter_filter(module, key, value):  # -> bool:
+        ...
+    def filter_and_map(
+        self,
+        filter_fn: Callable[[Module, str, Any], bool],
+        map_fn: Optional[Callable] = ...,
+        is_leaf_fn: Optional[Callable[[Module, str, Any], bool]] = ...,
+    ):  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
+        """Recursively filter the contents of the module using ``filter_fn``,
+        namely only select keys and values where ``filter_fn`` returns true.
+
+        This is used to implement :meth:`parameters` and :meth:`trainable_parameters`
+        but it can also be used to extract any subset of the module's parameters.
+
+        Args:
+            filter_fn (Callable): Given a value, the key in which it is found
+                and the containing module, decide whether to keep the value or
+                drop it.
+            map_fn (Callable, optional): Optionally transform the value before
+                returning it.
+            is_leaf_fn (Callable, optional): Given a value, the key in which it
+                is found and the containing module decide if it is a leaf.
+
+        Returns:
+            A dictionary containing the contents of the module recursively filtered
+        """
+
+    def parameters(
+        self,
+    ) -> mx.MX_ARRAY_TREE:
+        """Recursively return all the :class:`mlx.core.array` members of this Module
+        as a dict of dicts and lists."""
+
+    def trainable_parameters(
+        self,
+    ) -> mx.MX_ARRAY_TREE:  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
+        """Recursively return all the non frozen :class:`mlx.core.array` members of
+        this Module as a dict of dicts and lists."""
+
+    def children(
+        self,
+    ) -> mx.MX_ARRAY_TREE:  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
+        """Return the direct descendants of this Module instance."""
+
+    def leaf_modules(
+        self,
+    ) -> mx.MX_ARRAY_TREE:  # -> dict[Any, Any | dict[Any, Any | dict[Any, Any] | list[Any]] | dict[Any, Any] | list[Any]]:
+        """Return the submodules that do not contain other modules."""
+
+    def update(self, parameters: dict, strict: bool = ...) -> Module:
+        """Replace the parameters of this Module with the provided ones in the
+        dict of dicts and lists.
+
+        Commonly used by the optimizer to change the model to the updated
+        (optimized) parameters. Also used by the :meth:`value_and_grad` to set the
+        tracers in the model in order to compute gradients.
+
+        The passed in parameters dictionary need not be a full dictionary
+        similar to :meth:`parameters`. Only the provided locations will be
+        updated.
+
+        Args:
+            parameters (dict): A complete or partial dictionary of the modules
+                parameters.
+            strict (bool): If ``True`` checks that ``parameters`` is a
+                subset of the module's parameters. Default: ``True``.
+        Returns:
+            The module instance after updating the parameters.
+        """
+
+    def apply(
+        self,
+        map_fn: Callable[[mx.array], mx.array],
+        filter_fn: Optional[Callable[[Module, str, Any], bool]] = ...,
+    ) -> Module:
+        """Map all the parameters using the provided ``map_fn`` and immediately
+        update the module with the mapped parameters.
+
+        For instance running ``model.apply(lambda x: x.astype(mx.float16))``
+        casts all parameters to 16 bit floats.
+
+        Args:
+            map_fn (Callable): Maps an array to another array
+            filter_fn (Callable, optional): Filter to select which arrays to
+                map (default: :meth:`Module.valid_parameter_filter`).
+
+        Returns:
+            The module instance after updating the parameters.
+        """
+
+    def update_modules(self, modules: dict, strict: bool = ...) -> Module:
+        """Replace the child modules of this :class:`Module` instance with the
+        provided ones in the dict of dicts and lists.
+
+        It is the equivalent of :meth:`Module.update` but for modules instead
+        of parameters and allows us to flexibly edit complex architectures by
+        programmatically swapping layers.
+
+        The passed in parameters dictionary need not be a full dictionary
+        similar to :meth:`modules`. Only the provided locations will be
+        updated.
+
+        Args:
+            modules (dict): A complete or partial dictionary of the module's
+                submodules.
+            strict (bool): If ``True`` checks that ``modules`` is a
+                subset of the child modules of this instance. Default: ``True``.
+        Returns:
+            The module instance after updating the submodules.
+        """
+
+    def apply_to_modules(self, apply_fn: Callable[[str, Module], Any]) -> Module:
+        """Apply a function to all the modules in this instance (including this
+        instance).
+
+        Args:
+            apply_fn (Callable): The function to apply to the modules.
+
+        Returns:
+            The module instance after updating submodules.
+        """
+
+    def modules(self):  # -> list[Any]:
+        """Return a list with all the modules in this instance.
+
+        Returns:
+            A list of :class:`Module` instances.
+        """
+
+    def named_modules(self):  # -> list[Any]:
+        """Return a list with all the modules in this instance and their name
+        with dot notation.
+
+        Returns:
+            A list of tuples (str, :class:`Module`).
+        """
+
+    def freeze(
+        self,
+        *,
+        recurse: bool = ...,
+        keys: Optional[Union[str, List[str]]] = ...,
+        strict: bool = ...,
+    ) -> Module:
+        """Freeze the Module's parameters or some of them. Freezing a parameter means not
+        computing gradients for it.
+
+        This function is idempotent i.e. freezing a frozen model is a no-op.
+
+        Example:
+            For instance to only train the attention parameters from a Transformer:
+
+            .. code-block:: python
+
+                model = nn.Transformer()
+                model.freeze()
+                model.apply_to_modules(lambda k, v: v.unfreeze() if k.endswith("attention") else None)
+
+        Args:
+            recurse (bool, optional): If True then freeze the parameters of the
+                submodules as well. Default: ``True``.
+            keys (str or list[str], optional): If provided then only these
+                parameters will be frozen otherwise all the parameters of a
+                module. For instance freeze all biases by calling
+                ``module.freeze(keys="bias")``.
+            strict (bool, optional): If set to ``True`` validate that the passed keys exist.
+                Default: ``False``.
+
+        Returns:
+            The module instance after freezing the parameters.
+        """
+
+    def unfreeze(
+        self,
+        *,
+        recurse: bool = ...,
+        keys: Optional[Union[str, List[str]]] = ...,
+        strict: bool = ...,
+    ) -> Module:
+        """Unfreeze the Module's parameters or some of them.
+
+        This function is idempotent ie unfreezing a model that is not frozen is
+        a noop.
+
+        Example:
+
+            For instance to only train the biases of a Transformer one can do:
+
+            .. code-block:: python
+
+                model = nn.Transformer()
+                model.freeze()
+                model.unfreeze(keys="bias")
+
+        Args:
+            recurse (bool, optional): If True then unfreeze the parameters of the
+                submodules as well. Default: ``True``.
+            keys (str or list[str], optional): If provided then only these
+                parameters will be unfrozen otherwise all the parameters of a
+                module. For instance unfreeze all biases by calling
+                ``module.unfreeze(keys="bias")``.
+            strict (bool, optional): If set to ``True`` validate that the passed keys exist.
+                Default: ``False``.
+
+        Returns:
+            The module instance after unfreezing the parameters.
+        """
+
+    def train(self, mode: bool = ...) -> Module:
+        """Set the model in or out of training mode.
+
+        Training mode only applies to certain layers. For example
+        :obj:`Dropout` applies a random mask in training mode, but is the
+        identity in evaluation mode.
+
+        Args:
+            mode (bool): Indicate if the model should be in training or
+                evaluation mode. Default: ``True``.
+        Returns:
+            The module instance after updating the training mode.
+        """
+
+    def eval(self) -> Module:
+        """Set the model to evaluation mode.
+
+        See :func:`train`.
+        """
+
+    def set_dtype(
+        self, dtype: mx.Dtype, predicate: Optional[Callable[[mx.Dtype], bool]] = ...
+    ):  # -> None:
+        """Set the dtype of the module's parameters.
+
+        Args:
+            dtype (Dtype): The new dtype.
+            predicate (typing.Callable, optional): A predicate to select
+              parameters to cast. By default, only parameters of type
+              :attr:`floating` will be updated to avoid casting integer
+              parameters to the new dtype.
+        """

.mlx_typings/mlx/nn/layers/containers.pyi

@@ -0,0 +1,21 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Callable
+
+import mlx.core as mx
+from base import Module
+
+class Sequential(Module):
+    """A layer that calls the passed callables in order.
+
+    We can pass either modules or plain callables to the Sequential module. If
+    our functions have learnable parameters they should be implemented as
+    ``nn.Module`` instances.
+
+    Args:
+        modules (tuple of Callables): The modules to call in order
+    """
+    def __init__(self, *modules: Module | Callable[[mx.array], mx.array]) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/layers/convolution.pyi

@@ -0,0 +1,116 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Union
+
+import mlx.core as mx
+from base import Module
+
+class Conv1d(Module):
+    """Applies a 1-dimensional convolution over the multi-channel input sequence.
+
+    The channels are expected to be last i.e. the input shape should be ``NLC`` where:
+
+    * ``N`` is the batch dimension
+    * ``L`` is the sequence length
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels
+        out_channels (int): The number of output channels
+        kernel_size (int): The size of the convolution filters
+        stride (int, optional): The stride when applying the filter.
+            Default: ``1``.
+        padding (int, optional): How many positions to 0-pad the input with.
+            Default: ``0``.
+        dilation (int, optional): The dilation of the convolution.
+        groups (int, optional): The number of groups for the convolution.
+            Default: ``1``.
+        bias (bool, optional): If ``True`` add a learnable bias to the output.
+            Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = ...,
+        padding: int = ...,
+        dilation: int = ...,
+        groups: int = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Conv2d(Module):
+    """Applies a 2-dimensional convolution over the multi-channel input image.
+
+    The channels are expected to be last i.e. the input shape should be ``NHWC`` where:
+
+    * ``N`` is the batch dimension
+    * ``H`` is the input image height
+    * ``W`` is the input image width
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        kernel_size (int or tuple): The size of the convolution filters.
+        stride (int or tuple, optional): The size of the stride when
+            applying the filter. Default: ``1``.
+        padding (int or tuple, optional): How many positions to 0-pad
+            the input with. Default: ``0``.
+        dilation (int or tuple, optional): The dilation of the convolution.
+        groups (int, optional): The number of groups for the convolution.
+            Default: ``1``.
+        bias (bool, optional): If ``True`` add a learnable bias to the
+            output. Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, tuple],
+        stride: Union[int, tuple] = ...,
+        padding: Union[int, tuple] = ...,
+        dilation: Union[int, tuple] = ...,
+        groups: int = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x) -> mx.array: ...
+
+class Conv3d(Module):
+    """Applies a 3-dimensional convolution over the multi-channel input image.
+
+    The channels are expected to be last i.e. the input shape should be ``NDHWC`` where:
+
+    * ``N`` is the batch dimension
+    * ``D`` is the input image depth
+    * ``H`` is the input image height
+    * ``W`` is the input image width
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        kernel_size (int or tuple): The size of the convolution filters.
+        stride (int or tuple, optional): The size of the stride when
+            applying the filter. Default: ``1``.
+        dilation (int or tuple, optional): The dilation of the convolution.
+        padding (int or tuple, optional): How many positions to 0-pad
+            the input with. Default: ``0``.
+        bias (bool, optional): If ``True`` add a learnable bias to the
+            output. Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, tuple],
+        stride: Union[int, tuple] = ...,
+        padding: Union[int, tuple] = ...,
+        dilation: Union[int, tuple] = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/layers/convolution_transpose.pyi

@@ -0,0 +1,119 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Union
+
+import mlx.core as mx
+from base import Module
+
+class ConvTranspose1d(Module):
+    """Applies a 1-dimensional transposed convolution over the multi-channel input sequence.
+
+    The channels are expected to be last i.e. the input shape should be ``NLC`` where:
+
+    * ``N`` is the batch dimension
+    * ``L`` is the sequence length
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels
+        out_channels (int): The number of output channels
+        kernel_size (int): The size of the convolution filters
+        stride (int, optional): The stride when applying the filter.
+            Default: ``1``.
+        padding (int, optional): How many positions to 0-pad the input with.
+            Default: ``0``.
+        dilation (int, optional): The dilation of the convolution.
+        output_padding(int, optional): Additional size added to one side of the
+            output shape. Default: ``0``.
+        bias (bool, optional): If ``True`` add a learnable bias to the output.
+            Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = ...,
+        padding: int = ...,
+        dilation: int = ...,
+        output_padding: int = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class ConvTranspose2d(Module):
+    """Applies a 2-dimensional transposed convolution over the multi-channel input image.
+
+    The channels are expected to be last i.e. the input shape should be ``NHWC`` where:
+
+    * ``N`` is the batch dimension
+    * ``H`` is the input image height
+    * ``W`` is the input image width
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        kernel_size (int or tuple): The size of the convolution filters.
+        stride (int or tuple, optional): The size of the stride when
+            applying the filter. Default: ``1``.
+        padding (int or tuple, optional): How many positions to 0-pad
+            the input with. Default: ``0``.
+        dilation (int or tuple, optional): The dilation of the convolution.
+        output_padding(int or tuple, optional): Additional size added to one
+            side of the output shape. Default: ``0``.
+        bias (bool, optional): If ``True`` add a learnable bias to the
+            output. Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, tuple],
+        stride: Union[int, tuple] = ...,
+        padding: Union[int, tuple] = ...,
+        dilation: Union[int, tuple] = ...,
+        output_padding: Union[int, tuple] = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class ConvTranspose3d(Module):
+    """Applies a 3-dimensional transposed convolution over the multi-channel input image.
+
+    The channels are expected to be last i.e. the input shape should be ``NDHWC`` where:
+
+    * ``N`` is the batch dimension
+    * ``D`` is the input image depth
+    * ``H`` is the input image height
+    * ``W`` is the input image width
+    * ``C`` is the number of input channels
+
+    Args:
+        in_channels (int): The number of input channels.
+        out_channels (int): The number of output channels.
+        kernel_size (int or tuple): The size of the convolution filters.
+        stride (int or tuple, optional): The size of the stride when
+            applying the filter. Default: ``1``.
+        padding (int or tuple, optional): How many positions to 0-pad
+            the input with. Default: ``0``.
+        dilation (int or tuple, optional): The dilation of the convolution.
+        output_padding(int or tuple, optional): Additional size added to one
+            side of the output shape. Default: ``0``.
+        bias (bool, optional): If ``True`` add a learnable bias to the
+            output. Default: ``True``
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: Union[int, tuple],
+        stride: Union[int, tuple] = ...,
+        padding: Union[int, tuple] = ...,
+        dilation: Union[int, tuple] = ...,
+        output_padding: Union[int, tuple] = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/layers/distributed.pyi

@@ -0,0 +1,227 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from functools import lru_cache
+from typing import Callable, Optional, Union
+
+import mlx.core as mx
+from base import Module
+from mlx.nn.layers.linear import Linear
+
+@lru_cache
+def sum_gradients(
+    group: mx.distributed.Group,
+) -> Callable[..., mx.array]:  # -> Callable[..., Any] | Callable[..., array]:
+    ...
+def shard_inplace(
+    module: Module,
+    sharding: str,
+    *,
+    segments: Union[int, list[int]] = ...,
+    group: Optional[mx.distributed.Group] = ...,
+) -> None:
+    """Shard a module in-place by updating its parameter dictionary with the
+    sharded parameter dictionary.
+
+    The ``sharding`` argument can be any callable that given the path and the
+    weight returns the sharding axis and optionally also the segments that
+    comprise the unsharded weight. For instance if the weight is a fused QKV
+    matrix the segments should be 3.
+
+    .. note::
+        The module doesn't change so in order for distributed communication to
+        happen the module needs to natively support it and for it to be enabled.
+
+    Args:
+        module (Module): The parameters of this module will be sharded
+            in-place.
+        sharding (str or callable): One of "all-to-sharded" and
+            "sharded-to-all" or a callable that returns the sharding axis and
+            segments.
+        segments (int or list): The segments to use if ``sharding`` is a
+            string. Default: ``1``.
+        group (mlx.core.distributed.Group): The distributed group to shard
+            across. If not set, the global group will be used. Default: ``None``.
+    """
+
+def shard_linear(
+    module: Module,
+    sharding: str,
+    *,
+    segments: Union[int, list[int]] = ...,
+    group: Optional[mx.distributed.Group] = ...,
+) -> Linear:
+    """Create a new linear layer that has its parameters sharded and also
+    performs distributed communication either in the forward or backward
+    pass.
+
+    .. note::
+        Contrary to ``shard_inplace``, the original layer is not changed but a
+        new layer is returned.
+
+    Args:
+        module (Module): The linear layer to be sharded.
+        sharding (str): One of "all-to-sharded" and
+            "sharded-to-all" that defines the type of sharding to perform.
+        segments (int or list): The segments to use. Default: ``1``.
+        group (mlx.core.distributed.Group): The distributed group to shard
+            across. If not set, the global group will be used. Default: ``None``.
+    """
+
+class AllToShardedLinear(Module):
+    """Each member of the group applies part of the affine transformation such
+    that the result is sharded across the group.
+
+    The gradients are automatically aggregated from each member of the group.
+
+    Args:
+        input_dims (int): The dimensionality of the input features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` the the layer will not use a
+            bias. Default is ``True``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    @classmethod
+    def from_linear(
+        cls,
+        linear_layer: Module,
+        *,
+        segments: Union[int, list[int]] = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> AllToShardedLinear: ...
+
+class ShardedToAllLinear(Module):
+    """Each member of the group applies part of the affine transformation and
+    then aggregates the results.
+
+    All nodes will have the same exact result after this layer.
+
+    :class:`ShardedToAllLinear` provides a classmethod :meth:`from_linear` to
+    convert linear layers to sharded :obj:`ShardedToAllLinear` layers.
+
+    Args:
+        input_dims (int): The dimensionality of the input features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` the the layer will not use a
+            bias. Default is ``True``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    @classmethod
+    def from_linear(
+        cls,
+        linear_layer: Module,
+        *,
+        segments: Union[int, list[int]] = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> ShardedToAllLinear: ...
+
+class QuantizedAllToShardedLinear(Module):
+    """Each member of the group applies part of the affine transformation with
+    a quantized matrix such that the result is sharded across the group.
+
+    It is the quantized equivalent of :class:`AllToShardedLinear`.
+    Similar to :class:`QuantizedLinear` its parameters are frozen and
+    will not be included in any gradient computation.
+
+    Args:
+        input_dims (int): The dimensionality of the input features.
+        output_dims (int): The dimensionality of the output features.
+        bias (bool, optional): If set to ``False`` then the layer will not use
+            a bias. Default: ``True``.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = ...,
+        group_size: int = ...,
+        bits: int = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> None: ...
+    def unfreeze(self, *args, **kwargs) -> None:
+        """Wrap unfreeze so that we unfreeze any layers we might contain but
+        our parameters will remain frozen."""
+
+    def __call__(self, x: mx.array) -> mx.array: ...
+    @classmethod
+    def from_quantized_linear(
+        cls,
+        quantized_linear_layer: Module,
+        *,
+        segments: Union[int, list[int]] = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> QuantizedAllToShardedLinear: ...
+
+class QuantizedShardedToAllLinear(Module):
+    """Each member of the group applies part of the affine transformation using
+    the quantized matrix and then aggregates the results.
+
+    All nodes will have the same exact result after this layer.
+
+    It is the quantized equivalent of :class:`ShardedToAllLinear`.
+    Similar to :class:`QuantizedLinear` its parameters are frozen and
+    will not be included in any gradient computation.
+
+    Args:
+        input_dims (int): The dimensionality of the input features.
+        output_dims (int): The dimensionality of the output features.
+        bias (bool, optional): If set to ``False`` then the layer will not use
+            a bias. Default: ``True``.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = ...,
+        group_size: int = ...,
+        bits: int = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> None: ...
+    def unfreeze(self, *args, **kwargs):  # -> None:
+        """Wrap unfreeze so that we unfreeze any layers we might contain but
+        our parameters will remain frozen."""
+
+    def __call__(self, x: mx.array) -> mx.array: ...
+    @classmethod
+    def from_quantized_linear(
+        cls,
+        quantized_linear_layer: Module,
+        *,
+        segments: Union[int, list[int]] = ...,
+        group: Optional[mx.distributed.Group] = ...,
+    ) -> QuantizedShardedToAllLinear: ...

.mlx_typings/mlx/nn/layers/dropout.pyi

@@ -0,0 +1,65 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.core as mx
+from base import Module
+
+class Dropout(Module):
+    r"""Randomly zero a portion of the elements during training.
+
+    The remaining elements are multiplied with :math:`\frac{1}{1-p}` where
+    :math:`p` is the probability of zeroing an element. This is done so the
+    expected value of a given element will remain the same.
+
+    Args:
+        p (float): The probability to zero an element
+    """
+    def __init__(self, p: float = ...) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Dropout2d(Module):
+    r"""Apply 2D channel-wise dropout during training.
+
+    Randomly zero out entire channels independently with probability :math:`p`.
+    This layer expects the channels to be last, i.e. the input shape should be
+    ``NWHC`` or ``WHC`` where:``N`` is the batch dimension,``H`` is the input
+    image height,``W`` is the input image width, and``C`` is the number of
+    input channels
+
+    The remaining channels are scaled by :math:`\frac{1}{1-p}` to
+    maintain the expected value of each element. Unlike traditional dropout,
+    which zeros individual entries, this layer zeros entire channels. This is
+    beneficial for early convolution layers where adjacent pixels are
+    correlated. In such case, traditional dropout may not effectively
+    regularize activations. For more details, see [1].
+
+    [1]: Thompson, J., Goroshin, R., Jain, A., LeCun, Y. and Bregler C., 2015.
+    Efficient Object Localization Using Convolutional Networks. CVPR 2015.
+
+    Args:
+        p (float): Probability of zeroing a channel during training.
+    """
+    def __init__(self, p: float = ...) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Dropout3d(Module):
+    r"""Apply 3D channel-wise dropout during training.
+
+    Randomly zero out entire channels independently with probability :math:`p`.
+    This layer expects the channels to be last, i.e., the input shape should be
+    `NDHWC` or `DHWC` where: `N` is the batch dimension, `D` is the depth,
+    `H` is the input image height, `W` is the input image width, and `C` is
+    the number of input channels.
+
+    The remaining channels are scaled by :math:`\frac{1}{1-p}` to
+    maintain the expected value of each element. Unlike traditional dropout,
+    which zeros individual entries, this layer zeros entire channels. This is
+    often beneficial for convolutional layers processing 3D data, like in
+    medical imaging or video processing.
+
+    Args:
+        p (float): Probability of zeroing a channel during training.
+    """
+    def __init__(self, p: float = ...) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/layers/embedding.pyi

@@ -0,0 +1,34 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.core as mx
+from base import Module
+
+from .quantized import QuantizedEmbedding
+
+class Embedding(Module):
+    """Implements a simple lookup table that maps each input integer to a
+    high-dimensional vector.
+
+    Typically used to embed discrete tokens for processing by neural networks.
+
+    Args:
+        num_embeddings (int): How many possible discrete tokens can we embed.
+           Usually called the vocabulary size.
+        dims (int): The dimensionality of the embeddings.
+    """
+    def __init__(self, num_embeddings: int, dims: int) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    def as_linear(self, x: mx.array) -> mx.array:
+        """
+        Call the embedding layer as a linear layer.
+
+        Use this for example when input embedding and output projection
+        weights are tied.
+        """
+
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ..., mode: str = ...
+    ) -> QuantizedEmbedding:
+        """Return a :obj:`QuantizedEmbedding` layer that approximates this embedding layer."""

.mlx_typings/mlx/nn/layers/linear.pyi

@@ -0,0 +1,76 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any
+
+import mlx.core as mx
+from base import Module
+
+from .quantized import QuantizedLinear
+
+class Identity(Module):
+    r"""A placeholder identity operator that is argument-insensitive.
+
+    Args:
+        args: any argument (unused)
+        kwargs: any keyword argument (unused)
+    """
+    def __init__(self, *args: Any, **kwargs: Any) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class Linear(Module):
+    r"""Applies an affine transformation to the input.
+
+    Concretely:
+
+    .. math::
+
+        y = x W^\top + b
+
+    where:
+    where :math:`W` has shape ``[output_dims, input_dims]`` and :math:`b` has shape ``[output_dims]``.
+
+    The values are initialized from the uniform distribution :math:`\mathcal{U}(-{k}, {k})`,
+    where :math:`k = \frac{1}{\sqrt{D_i}}` and :math:`D_i` is equal to ``input_dims``.
+
+    Args:
+        input_dims (int): The dimensionality of the input features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` then the layer will
+          not use a bias. Default is ``True``.
+    """
+    def __init__(self, input_dims: int, output_dims: int, bias: bool = ...) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ..., mode: str = ...
+    ) -> QuantizedLinear:
+        """Return a :obj:`QuantizedLinear` layer that approximates this layer."""
+
+class Bilinear(Module):
+    r"""Applies a bilinear transformation to the inputs.
+
+    Concretely:
+
+    .. math::
+
+        y_i = x_1^\top W_i x_2 + b_i
+
+    where:
+    :math:`W` has shape ``[output_dims, input1_dims, input2_dims]``, :math:`b` has shape ``[output_dims ]``,
+    and :math:`i` indexes the output dimension.
+
+    The values are initialized from the uniform distribution :math:`\mathcal{U}(-{k}, {k})`,
+    where :math:`k = \frac{1}{\sqrt{D_1}}` and :math:`D_1` is ``input1_dims``.
+
+    Args:
+        input1_dims (int): The dimensionality of the input1 features
+        input2_dims (int): The dimensionality of the input2 features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` then the layer will
+          not use a bias. Default is ``True``.
+    """
+    def __init__(
+        self, input1_dims: int, input2_dims: int, output_dims: int, bias: bool = ...
+    ) -> None: ...
+    def __call__(self, x1: mx.array, x2: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/layers/normalization.pyi

@@ -0,0 +1,194 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.core as mx
+from base import Module
+
+class InstanceNorm(Module):
+    r"""Applies instance normalization [1] on the inputs.
+
+    Computes
+
+    .. math::
+
+        y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta,
+
+    where :math:`\gamma` and :math:`\beta` are learned per feature dimension
+    parameters initialized at 1 and 0 respectively. Both are of size :attr:`dims`,
+    if :attr:`affine` is ``True``.
+
+    Args:
+        dims (int): The number of features of the input.
+        eps (float): A value added to the denominator for numerical stability. Default: ``1e-5``.
+        affine (bool): Default: ``False``.
+
+    Shape:
+      - Input: :math:`(..., C)` where :math:`C` is equal to :attr:`dims`.
+      - Output: Same shape as the input.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+        >>> x = mx.random.normal((8, 4, 4, 16))
+        >>> inorm = nn.InstanceNorm(dims=16)
+        >>> output = inorm(x)
+
+    References:
+        [1]: https://arxiv.org/abs/1607.08022
+    """
+    def __init__(self, dims: int, eps: float = ..., affine: bool = ...) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class LayerNorm(Module):
+    r"""Applies layer normalization [1] on the inputs.
+
+    Computes
+
+    .. math::
+
+        y = \frac{x - E[x]}{\sqrt{Var[x]} + \epsilon} \gamma + \beta,
+
+    where :math:`\gamma` and :math:`\beta` are learned per feature dimension
+    parameters initialized at 1 and 0 respectively.
+
+    [1]: https://arxiv.org/abs/1607.06450
+
+    Args:
+        dims (int): The feature dimension of the input to normalize over
+        eps (float): A small additive constant for numerical stability
+        affine (bool): If True learn an affine transform to apply after the
+            normalization
+        bias (bool): If True include a translation to the affine
+            transformation. If set to False the transformation is not really affine
+            just scaling.
+    """
+    def __init__(
+        self, dims: int, eps: float = ..., affine: bool = ..., bias: bool = ...
+    ) -> None: ...
+    def __call__(self, x) -> mx.array: ...
+
+class RMSNorm(Module):
+    r"""Applies Root Mean Square normalization [1] to the inputs.
+
+    Computes
+
+    ..  math::
+
+        y = \frac{x}{\sqrt{E[x^2] + \epsilon}} \gamma
+
+    where :math:`\gamma` is a learned per feature dimension parameter initialized at
+    1.
+
+    Note the accumulation for the mean is done in 32-bit precision.
+
+    [1]: https://arxiv.org/abs/1910.07467
+
+    Args:
+        dims (int): The feature dimension of the input to normalize over
+        eps (float): A small additive constant for numerical stability
+    """
+    def __init__(self, dims: int, eps: float = ...) -> None: ...
+    def __call__(self, x) -> mx.array: ...
+
+class GroupNorm(Module):
+    r"""Applies Group Normalization [1] to the inputs.
+
+    Computes the same normalization as layer norm, namely
+
+    .. math::
+
+        y = \frac{x - E[x]}{\sqrt{Var[x]} + \epsilon} \gamma + \beta,
+
+    where :math:`\gamma` and :math:`\beta` are learned per feature dimension
+    parameters initialized at 1 and 0 respectively. However, the mean and
+    variance are computed over the spatial dimensions and each group of
+    features. In particular, the input is split into num_groups across the
+    feature dimension.
+
+    The feature dimension is assumed to be the last dimension and the dimensions
+    that precede it (except the first) are considered the spatial dimensions.
+
+    [1]: https://arxiv.org/abs/1803.08494
+
+    Args:
+        num_groups (int): Number of groups to separate the features into
+        dims (int): The feature dimensions of the input to normalize over
+        eps (float): A small additive constant for numerical stability
+        affine (bool): If True learn an affine transform to apply after the
+            normalization.
+        pytorch_compatible (bool): If True perform the group normalization in
+            the same order/grouping as PyTorch.
+    """
+    def __init__(
+        self,
+        num_groups: int,
+        dims: int,
+        eps: float = ...,
+        affine: bool = ...,
+        pytorch_compatible: bool = ...,
+    ) -> None: ...
+    def __call__(self, x) -> mx.array: ...
+
+class BatchNorm(Module):
+    r"""Applies Batch Normalization over a 2D or 3D input.
+
+    Computes
+
+    .. math::
+
+        y = \frac{x - E[x]}{\sqrt{Var[x]} + \epsilon} \gamma + \beta,
+
+    where :math:`\gamma` and :math:`\beta` are learned per feature dimension
+    parameters initialized at 1 and 0 respectively.
+
+    The input shape is specified as ``NC`` or ``NLC``, where ``N`` is the
+    batch, ``C`` is the number of features or channels, and ``L`` is the
+    sequence length. The output has the same shape as the input. For
+    four-dimensional arrays, the shape is ``NHWC``, where ``H`` and ``W`` are
+    the height and width respectively.
+
+    For more information on Batch Normalization, see the original paper `Batch
+    Normalization: Accelerating Deep Network Training by Reducing Internal
+    Covariate Shift <https://arxiv.org/abs/1502.03167>`_.
+
+    Args:
+        num_features (int): The feature dimension to normalize over.
+        eps (float, optional): A small additive constant for numerical
+            stability. Default: ``1e-5``.
+        momentum (float, optional): The momentum for updating the running
+            mean and variance. Default: ``0.1``.
+        affine (bool, optional): If ``True``, apply a learned affine
+            transformation after the normalization. Default: ``True``.
+        track_running_stats (bool, optional): If ``True``, track the
+            running mean and variance. Default: ``True``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+        >>> x = mx.random.normal((5, 4))
+        >>> bn = nn.BatchNorm(num_features=4, affine=True)
+        >>> output = bn(x)
+    """
+    def __init__(
+        self,
+        num_features: int,
+        eps: float = ...,
+        momentum: float = ...,
+        affine: bool = ...,
+        track_running_stats: bool = ...,
+    ) -> None: ...
+    def unfreeze(self, *args, **kwargs):  # -> None:
+        """Wrap unfreeze to make sure that running_mean and var are always
+        frozen parameters."""
+
+    def __call__(self, x: mx.array) -> mx.array:
+        """
+        Forward pass of BatchNorm.
+
+        Args:
+            x (array): Input tensor.
+
+        Returns:
+            array: Normalized output tensor.
+        """

.mlx_typings/mlx/nn/layers/pooling.pyi

@@ -0,0 +1,242 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Optional, Tuple, Union
+
+import mlx.core as mx
+from base import Module
+
+class _Pool(Module):
+    def __init__(
+        self, pooling_function, kernel_size, stride, padding, padding_value
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class _Pool1d(_Pool):
+    def __init__(
+        self,
+        pooling_function,
+        padding_value,
+        kernel_size: Union[int, Tuple[int]],
+        stride: Optional[Union[int, Tuple[int]]] = ...,
+        padding: Union[int, Tuple[int]] = ...,
+    ) -> None: ...
+
+class _Pool2d(_Pool):
+    def __init__(
+        self,
+        pooling_function,
+        padding_value,
+        kernel_size: Union[int, Tuple[int, int]],
+        stride: Optional[Union[int, Tuple[int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int]]] = ...,
+    ) -> None: ...
+
+class _Pool3d(_Pool):
+    def __init__(
+        self,
+        pooling_function,
+        padding_value,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Optional[Union[int, Tuple[int, int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int, int]]] = ...,
+    ) -> None: ...
+
+class MaxPool1d(_Pool1d):
+    r"""Applies 1-dimensional max pooling.
+
+    Spatially downsamples the input by taking the maximum of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    Args:
+        kernel_size (int or tuple(int)): The size of the pooling window kernel.
+        stride (int or tuple(int), optional): The stride of the pooling window.
+            Default: ``kernel_size``.
+        padding (int or tuple(int), optional): How much negative infinity
+            padding to apply to the input. The padding amount is applied to
+            both sides of the spatial axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(4, 16, 5))
+        >>> pool = nn.MaxPool1d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int]],
+        stride: Optional[Union[int, Tuple[int]]] = ...,
+        padding: Union[int, Tuple[int]] = ...,
+    ) -> None: ...
+
+class AvgPool1d(_Pool1d):
+    r"""Applies 1-dimensional average pooling.
+
+    Spatially downsamples the input by taking the average of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    Args:
+        kernel_size (int or tuple(int)): The size of the pooling window kernel.
+        stride (int or tuple(int), optional): The stride of the pooling window.
+            Default: ``kernel_size``.
+        padding (int or tuple(int), optional): How much zero padding to apply to
+            the input. The padding amount is applied to both sides of the spatial
+            axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(4, 16, 5))
+        >>> pool = nn.AvgPool1d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int]],
+        stride: Optional[Union[int, Tuple[int]]] = ...,
+        padding: Union[int, Tuple[int]] = ...,
+    ) -> None: ...
+
+class MaxPool2d(_Pool2d):
+    r"""Applies 2-dimensional max pooling.
+
+    Spatially downsamples the input by taking the maximum of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    The parameters ``kernel_size``, ``stride``, and ``padding`` can either be:
+
+    * a single ``int`` -- in which case the same value is used for both the
+      height and width axis.
+    * a ``tuple`` of two ``int`` s -- in which case, the first ``int`` is
+      used for the height axis, the second ``int`` for the width axis.
+
+    Args:
+        kernel_size (int or tuple(int, int)): The size of the pooling window.
+        stride (int or tuple(int, int), optional): The stride of the pooling
+            window. Default: ``kernel_size``.
+        padding (int or tuple(int, int), optional): How much negative infinity
+            padding to apply to the input. The padding is applied on both sides
+            of the height and width axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(8, 32, 32, 4))
+        >>> pool = nn.MaxPool2d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int, int]],
+        stride: Optional[Union[int, Tuple[int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int]]] = ...,
+    ) -> None: ...
+
+class AvgPool2d(_Pool2d):
+    r"""Applies 2-dimensional average pooling.
+
+    Spatially downsamples the input by taking the average of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    The parameters ``kernel_size``, ``stride``, and ``padding`` can either be:
+
+    * a single ``int`` -- in which case the same value is used for both the
+      height and width axis.
+    * a ``tuple`` of two ``int`` s -- in which case, the first ``int`` is
+      used for the height axis, the second ``int`` for the width axis.
+
+    Args:
+        kernel_size (int or tuple(int, int)): The size of the pooling window.
+        stride (int or tuple(int, int), optional): The stride of the pooling
+            window. Default: ``kernel_size``.
+        padding (int or tuple(int, int), optional): How much zero
+            padding to apply to the input. The padding is applied on both sides
+            of the height and width axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(8, 32, 32, 4))
+        >>> pool = nn.AvgPool2d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int, int]],
+        stride: Optional[Union[int, Tuple[int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int]]] = ...,
+    ) -> None: ...
+
+class MaxPool3d(_Pool3d):
+    r"""Applies 3-dimensional max pooling.
+
+    Spatially downsamples the input by taking the maximum of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    The parameters ``kernel_size``, ``stride``, and ``padding`` can either be:
+
+    * a single ``int`` -- in which case the same value is used for the depth,
+      height, and width axis.
+    * a ``tuple`` of three ``int`` s -- in which case, the first ``int`` is used
+      for the depth axis, the second ``int`` for the height axis, and the third
+      ``int`` for the width axis.
+
+    Args:
+        kernel_size (int or tuple(int, int, int)): The size of the pooling window.
+        stride (int or tuple(int, int, int), optional): The stride of the pooling
+            window. Default: ``kernel_size``.
+        padding (int or tuple(int, int, int), optional): How much negative infinity
+            padding to apply to the input. The padding is applied on both sides
+            of the depth, height and width axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(8, 16, 32, 32, 4))
+        >>> pool = nn.MaxPool3d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Optional[Union[int, Tuple[int, int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int, int]]] = ...,
+    ) -> None: ...
+
+class AvgPool3d(_Pool3d):
+    r"""Applies 3-dimensional average pooling.
+
+    Spatially downsamples the input by taking the average of a sliding window
+    of size ``kernel_size`` and sliding stride ``stride``.
+
+    The parameters ``kernel_size``, ``stride``, and ``padding`` can either be:
+
+    * a single ``int`` -- in which case the same value is used for the depth,
+      height, and width axis.
+    * a ``tuple`` of three ``int`` s -- in which case, the first ``int`` is used
+      for the depth axis, the second ``int`` for the height axis, and the third
+      ``int`` for the width axis.
+
+    Args:
+        kernel_size (int or tuple(int, int, int)): The size of the pooling window.
+        stride (int or tuple(int, int, int), optional): The stride of the pooling
+            window. Default: ``kernel_size``.
+        padding (int or tuple(int, int, int), optional): How much zero
+            padding to apply to the input. The padding is applied on both sides
+            of the depth, height and width axis. Default: ``0``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import layers as nn
+        >>> x = mx.random.normal(shape=(8, 16, 32, 32, 4))
+        >>> pool = nn.AvgPool3d(kernel_size=2, stride=2)
+        >>> pool(x)
+    """
+    def __init__(
+        self,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Optional[Union[int, Tuple[int, int, int]]] = ...,
+        padding: Optional[Union[int, Tuple[int, int, int]]] = ...,
+    ) -> None: ...

.mlx_typings/mlx/nn/layers/positional_encoding.pyi

@@ -0,0 +1,80 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Optional
+
+import mlx.core as mx
+from base import Module
+
+class RoPE(Module):
+    """Implements the rotary positional encoding.
+
+    The traditional implementation rotates consecutive pairs of elements in the
+    feature dimension while the default implementation rotates pairs with
+    stride half the feature dimensions for efficiency.
+
+    For more details see `RoFormer: Enhanced Transformer with Rotary Position
+    Embedding <https://arxiv.org/abs/2104.09864>`_.
+
+    Args:
+        dims (int): The feature dimensions to be rotated. If the input feature
+            is larger than dims then the rest is left unchanged.
+        traditional (bool, optional): If set to ``True`` choose the traditional
+            implementation which is slightly less efficient. Default: ``False``.
+        base (float, optional): The base used to compute angular frequency for
+            each dimension in the positional encodings. Default: ``10000``.
+        scale (float, optional): The scale used to scale the positions. Default: ``1.0``.
+    """
+    def __init__(
+        self, dims: int, traditional: bool = ..., base: float = ..., scale: float = ...
+    ) -> None: ...
+    def __call__(self, x, offset: int = ...) -> mx.array: ...
+
+class SinusoidalPositionalEncoding(Module):
+    r"""Implements sinusoidal positional encoding.
+
+    For more details see the paper `Attention Is All You Need
+    <https://arxiv.org/abs/1706.03762>`_.
+
+    Args:
+        dims (int): The dimensionality of the resulting positional embeddings.
+        min_freq (float, optional): The minimum frequency expected. Default:
+            ``0.0001``.
+        max_freq (float, optional): The maximum frequency expected. Default:
+            ``1``.
+        scale (float, optional): A multiplicative scale for the embeddings.
+            Default: ``sqrt(2/dims)``.
+        cos_first (bool, optional): If ``True`` embed using ``[cos(x); sin(x)]``
+            instead of the reverse. Default: ``False``.
+        full_turns (bool, optional): If ``True`` multiply the frequencies with
+            :math:`2\pi`. Default: ``False``.
+    """
+    def __init__(
+        self,
+        dims: int,
+        min_freq: float = ...,
+        max_freq: float = ...,
+        scale: Optional[float] = ...,
+        cos_first: bool = ...,
+        full_turns: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+
+class ALiBi(Module):
+    _alibi_mask_key = ...
+    _alibi_mask = ...
+    @classmethod
+    def create_alibi_matrix(
+        cls,
+        q_sequence_length: int,
+        k_sequence_length: int,
+        num_heads: int,
+        offset: int,
+        dtype=...,
+    ) -> mx.array | None: ...
+    @staticmethod
+    def create_alibi_slope(num_heads: int) -> mx.array: ...
+    def __call__(
+        self, attention_scores: mx.array, offset=..., mask=...
+    ) -> mx.array: ...

.mlx_typings/mlx/nn/layers/quantized.pyi

@@ -0,0 +1,125 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Callable, Optional, Union
+
+import mlx.core as mx
+from base import Module
+
+def quantize(
+    model: Module,
+    group_size: int = ...,
+    bits: int = ...,
+    *,
+    mode: str = ...,
+    class_predicate: Optional[Callable[[str, Module], Union[bool, dict]]] = ...,
+):  # -> None:
+    """Quantize the sub-modules of a module according to a predicate.
+
+    By default all layers that define a ``to_quantized(group_size, bits)``
+    method will be quantized. Both :obj:`Linear` and :obj:`Embedding` layers
+    will be quantized. Note also, the module is updated in-place.
+
+    Args:
+        model (Module): The model whose leaf modules may be quantized.
+        group_size (int): The quantization group size (see
+           :func:`mlx.core.quantize`). Default: ``64``.
+        bits (int): The number of bits per parameter (see
+           :func:`mlx.core.quantize`). Default: ``4``.
+        mode (str): The quantization method to use (see
+           :func:`mlx.core.quantize`). Default: ``"affine"``.
+        class_predicate (Optional[Callable]): A callable which receives the
+          :obj:`Module` path and :obj:`Module` itself and returns ``True`` or a
+          dict of params for `to_quantized` if it should be quantized and
+          ``False`` otherwise. If ``None``, then all layers that define a
+          ``to_quantized(group_size, bits)`` method are quantized.
+          Default: ``None``.
+    """
+
+class QuantizedEmbedding(Module):
+    """The same as :obj:`Embedding` but with a  quantized weight matrix.
+
+    :obj:`QuantizedEmbedding` also provides a :meth:`from_embedding`
+    classmethod to convert embedding layers to :obj:`QuantizedEmbedding`
+    layers.
+
+    Args:
+        num_embeddings (int): How many possible discrete tokens can we embed.
+           Usually called the vocabulary size.
+        dims (int): The dimensionality of the embeddings.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        mode (str): The quantization method to use (see
+           :func:`mlx.core.quantize`). Default: ``"affine"``.
+    """
+    def __init__(
+        self,
+        num_embeddings: int,
+        dims: int,
+        group_size: int = ...,
+        bits: int = ...,
+        mode: str = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    def as_linear(self, x: mx.array) -> mx.array:
+        """
+        Call the quantized embedding layer as a quantized linear layer.
+
+        Use this for example when input embedding and output projection
+        weights are tied.
+        """
+
+    @classmethod
+    def from_embedding(
+        cls,
+        embedding_layer: Module,
+        group_size: int = ...,
+        bits: int = ...,
+        mode: str = ...,
+    ) -> QuantizedEmbedding:
+        """Create a :obj:`QuantizedEmbedding` layer from an :obj:`Embedding` layer."""
+
+class QuantizedLinear(Module):
+    """Applies an affine transformation to the input using a quantized weight matrix.
+
+    It is the quantized equivalent of :class:`Linear`. For now its
+    parameters are frozen and will not be included in any gradient computation
+    but this will probably change in the future.
+
+    :obj:`QuantizedLinear` also provides a classmethod :meth:`from_linear` to
+    convert linear layers to :obj:`QuantizedLinear` layers.
+
+    Args:
+        input_dims (int): The dimensionality of the input features.
+        output_dims (int): The dimensionality of the output features.
+        bias (bool, optional): If set to ``False`` then the layer will not use
+            a bias. Default: ``True``.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        mode (str): The quantization method to use (see
+           :func:`mlx.core.quantize`). Default: ``"affine"``.
+    """
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = ...,
+        group_size: int = ...,
+        bits: int = ...,
+        mode: str = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...
+    @classmethod
+    def from_linear(
+        cls,
+        linear_layer: Module,
+        group_size: int = ...,
+        bits: int = ...,
+        mode: str = ...,
+    ) -> QuantizedLinear:
+        """Create a :obj:`QuantizedLinear` layer from a :obj:`Linear` layer."""

.mlx_typings/mlx/nn/layers/recurrent.pyi

@@ -0,0 +1,113 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Callable, Optional
+
+import mlx.core as mx
+from base import Module
+
+class RNN(Module):
+    r"""An Elman recurrent layer.
+
+    The input is a sequence of shape ``NLD`` or ``LD`` where:
+
+    * ``N`` is the optional batch dimension
+    * ``L`` is the sequence length
+    * ``D`` is the input's feature dimension
+
+    Concretely, for each element along the sequence length axis, this
+    layer applies the function:
+
+    .. math::
+
+        h_{t + 1} = \text{tanh} (W_{ih}x_t + W_{hh}h_t + b)
+
+    The hidden state :math:`h` has shape ``NH`` or ``H``, depending on
+    whether the input is batched or not. Returns the hidden state at each
+    time step, of shape ``NLH`` or ``LH``.
+
+    Args:
+        input_size (int): Dimension of the input, ``D``.
+        hidden_size (int): Dimension of the hidden state, ``H``.
+        bias (bool, optional): Whether to use a bias. Default: ``True``.
+        nonlinearity (callable, optional): Non-linearity to use. If ``None``,
+            then func:`tanh` is used. Default: ``None``.
+    """
+    def __init__(
+        self,
+        input_size: int,
+        hidden_size: int,
+        bias: bool = ...,
+        nonlinearity: Optional[Callable] = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array, hidden=...) -> mx.array: ...
+
+class GRU(Module):
+    r"""A gated recurrent unit (GRU) RNN layer.
+
+    The input has shape ``NLD`` or ``LD`` where:
+
+    * ``N`` is the optional batch dimension
+    * ``L`` is the sequence length
+    * ``D`` is the input's feature dimension
+
+    Concretely, for each element of the sequence, this layer computes:
+
+    .. math::
+
+        \begin{aligned}
+        r_t &= \sigma (W_{xr}x_t + W_{hr}h_t + b_{r}) \\
+        z_t &= \sigma (W_{xz}x_t + W_{hz}h_t + b_{z}) \\
+        n_t &= \text{tanh}(W_{xn}x_t + b_{n} + r_t \odot (W_{hn}h_t + b_{hn})) \\
+        h_{t + 1} &= (1 - z_t) \odot n_t + z_t \odot h_t
+        \end{aligned}
+
+    The hidden state :math:`h` has shape ``NH`` or ``H`` depending on
+    whether the input is batched or not. Returns the hidden state at each
+    time step of shape ``NLH`` or ``LH``.
+
+    Args:
+        input_size (int): Dimension of the input, ``D``.
+        hidden_size (int): Dimension of the hidden state, ``H``.
+        bias (bool): Whether to use biases or not. Default: ``True``.
+    """
+    def __init__(self, input_size: int, hidden_size: int, bias: bool = ...) -> None: ...
+    def __call__(self, x: mx.array, hidden=...) -> mx.array: ...
+
+class LSTM(Module):
+    r"""An LSTM recurrent layer.
+
+    The input has shape ``NLD`` or ``LD`` where:
+
+    * ``N`` is the optional batch dimension
+    * ``L`` is the sequence length
+    * ``D`` is the input's feature dimension
+
+    Concretely, for each element of the sequence, this layer computes:
+
+    .. math::
+        \begin{aligned}
+        i_t &= \sigma (W_{xi}x_t + W_{hi}h_t + b_{i}) \\
+        f_t &= \sigma (W_{xf}x_t + W_{hf}h_t + b_{f}) \\
+        g_t &= \text{tanh} (W_{xg}x_t + W_{hg}h_t + b_{g}) \\
+        o_t &= \sigma (W_{xo}x_t + W_{ho}h_t + b_{o}) \\
+        c_{t + 1} &= f_t \odot c_t + i_t \odot g_t \\
+        h_{t + 1} &= o_t \text{tanh}(c_{t + 1})
+        \end{aligned}
+
+    The hidden state :math:`h` and cell state :math:`c` have shape ``NH``
+    or ``H``, depending on whether the input is batched or not.
+
+    The layer returns two arrays, the hidden state and the cell state at
+    each time step, both of shape ``NLH`` or ``LH``.
+
+    Args:
+        input_size (int): Dimension of the input, ``D``.
+        hidden_size (int): Dimension of the hidden state, ``H``.
+        bias (bool): Whether to use biases or not. Default: ``True``.
+    """
+    def __init__(self, input_size: int, hidden_size: int, bias: bool = ...) -> None: ...
+    def __call__(
+        self, x: mx.array, hidden=..., cell=...
+    ) -> tuple[mx.array, mx.array]: ...

.mlx_typings/mlx/nn/layers/transformer.pyi

@@ -0,0 +1,168 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any, Callable, Optional
+
+import mlx.core as mx
+from base import Module
+
+class MultiHeadAttention(Module):
+    """Implements the scaled dot product attention with multiple heads.
+
+    Given inputs for queries, keys and values the ``MultiHeadAttention``
+    produces new values by aggregating information from the input values
+    according to the similarities of the input queries and keys.
+
+    All inputs as well as the output are linearly projected without biases by
+    default.
+
+    ``MultiHeadAttention`` also takes an optional additive attention mask that
+    should be broadcastable with ``(batch, num_heads, # queries, # keys)``. The
+    mask should have ``-inf`` or very large negative numbers at the positions
+    that should *not* be attended to.
+
+    Args:
+        dims (int): The model dimensions. This is also the default
+            value for the queries, keys, values, and the output.
+        num_heads (int): The number of attention heads to use.
+        query_input_dims (int, optional): The input dimensions of the queries.
+            Default: ``dims``.
+        key_input_dims (int, optional): The input dimensions of the keys.
+            Default: ``dims``.
+        value_input_dims (int, optional): The input dimensions of the values.
+            Default: ``key_input_dims``.
+        value_dims (int, optional): The dimensions of the values after the
+            projection. Default: ``dims``.
+        value_output_dims (int, optional): The dimensions the new values will
+            be projected to. Default: ``dims``.
+        bias (bool, optional): Whether or not to use a bias in the projections.
+            Default: ``False``.
+    """
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        query_input_dims: Optional[int] = ...,
+        key_input_dims: Optional[int] = ...,
+        value_input_dims: Optional[int] = ...,
+        value_dims: Optional[int] = ...,
+        value_output_dims: Optional[int] = ...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(
+        self, queries: mx.array, keys: mx.array, values: mx.array, mask: mx.array = ...
+    ) -> mx.array: ...
+    @staticmethod
+    def create_additive_causal_mask(N: int, dtype: mx.Dtype = ...) -> mx.array: ...
+
+class TransformerEncoderLayer(Module):
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = ...,
+        dropout: float = ...,
+        activation: Callable[[Any], Any] = ...,
+        norm_first: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array, mask: mx.array) -> mx.array: ...
+
+class TransformerEncoder(Module):
+    def __init__(
+        self,
+        num_layers: int,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = ...,
+        dropout: float = ...,
+        activation=...,
+        norm_first: bool = ...,
+        checkpoint: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array, mask: mx.array) -> mx.array: ...
+
+class TransformerDecoderLayer(Module):
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = ...,
+        dropout: float = ...,
+        activation: Callable[[Any], Any] = ...,
+        norm_first: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array, memory, x_mask, memory_mask) -> mx.array: ...
+
+class TransformerDecoder(Module):
+    def __init__(
+        self,
+        num_layers: int,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = ...,
+        dropout: float = ...,
+        activation=...,
+        norm_first: bool = ...,
+        checkpoint: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array, memory, x_mask, memory_mask) -> mx.array: ...
+
+class Transformer(Module):
+    """
+    Implements a standard Transformer model.
+
+    The implementation is based on `Attention Is All You Need
+    <https://arxiv.org/abs/1706.03762>`_.
+
+    The Transformer model contains an encoder and a decoder. The encoder
+    processes the input sequence and the decoder generates the output sequence.
+    The interaction between encoder and decoder happens through the attention
+    mechanism.
+
+    Args:
+        dims (int, optional): The number of expected features in the
+            encoder/decoder inputs. Default: ``512``.
+        num_heads (int, optional): The number of attention heads. Default:
+            ``8``.
+        num_encoder_layers (int, optional): The number of encoder layers in the
+            Transformer encoder. Default: ``6``.
+        num_decoder_layers (int, optional): The number of decoder layers in the
+            Transformer decoder. Default: ``6``.
+        mlp_dims (int, optional): The hidden dimension of the MLP block in each
+            Transformer layer. Defaults to ``4*dims`` if not provided. Default:
+            ``None``.
+        dropout (float, optional): The dropout value for the Transformer
+            encoder and decoder. Dropout is used after each attention layer and
+            the activation in the MLP layer. Default: ``0.0``.
+        activation (function, optional): the activation function for the MLP
+            hidden layer. Default: :func:`relu`.
+        custom_encoder (nn.Module, optional): A custom encoder to replace the
+            standard Transformer encoder. Default: ``None``.
+        custom_decoder (nn.Module, optional): A custom decoder to replace the
+            standard Transformer decoder. Default: ``None``.
+        norm_first (bool, optional): if ``True``, encoder and decoder layers
+            will perform layer normalization before attention and MLP
+            operations, otherwise after. Default: ``True``.
+        checkpoint (bool, optional): if ``True`` perform gradient checkpointing
+            to reduce the memory usage at the expense of more computation.
+            Default: ``False``.
+    """
+    def __init__(
+        self,
+        dims: int = ...,
+        num_heads: int = ...,
+        num_encoder_layers: int = ...,
+        num_decoder_layers: int = ...,
+        mlp_dims: Optional[int] = ...,
+        dropout: float = ...,
+        activation: Callable[[Any], Any] = ...,
+        custom_encoder: Optional[Any] = ...,
+        custom_decoder: Optional[Any] = ...,
+        norm_first: bool = ...,
+        checkpoint: bool = ...,
+    ) -> None: ...
+    def __call__(
+        self, src, tgt, src_mask, tgt_mask, memory_mask
+    ) -> mx.array:  # -> array | Any:
+        ...

.mlx_typings/mlx/nn/layers/upsample.pyi

@@ -0,0 +1,87 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Literal, Tuple, Union
+
+import mlx.core as mx
+from base import Module
+
+def upsample_nearest(x: mx.array, scale_factor: Tuple) -> mx.array: ...
+def upsample_linear(
+    x: mx.array, scale_factor: Tuple, align_corners: bool = ...
+):  # -> int:
+    ...
+def upsample_cubic(
+    x: mx.array, scale_factor: Tuple, align_corners: bool = ...
+):  # -> int:
+    ...
+
+class Upsample(Module):
+    r"""Upsample the input signal spatially.
+
+    The spatial dimensions are by convention dimensions ``1`` to ``x.ndim -
+    2``. The first is the batch dimension and the last is the feature
+    dimension.
+
+    For example, an audio signal would be 3D with 1 spatial dimension, an image
+    4D with 2 and so on and so forth.
+
+    There are three upsampling algorithms implemented nearest neighbor upsampling,
+    linear interpolation, and cubic interpolation. All can be applied to any number
+    of spatial dimensions. The linear interpolation will be bilinear, trilinear etc
+    when applied to more than one spatial dimension. And cubic interpolation will be
+    bicubic when there are 2 spatial dimensions.
+
+    .. note::
+       When using one of the linear or cubic interpolation modes the ``align_corners``
+       argument changes how the corners are treated in the input image. If
+       ``align_corners=True`` then the top and left edge of the input and
+       output will be matching as will the bottom right edge.
+
+    Parameters:
+        scale_factor (float or tuple): The multiplier for the spatial size.
+            If a ``float`` is provided, it is the multiplier for all spatial dimensions.
+            Otherwise, the number of scale factors provided must match the
+            number of spatial dimensions.
+        mode (str, optional): The upsampling algorithm, either ``"nearest"``,
+            ``"linear"`` or ``"cubic"``. Default: ``"nearest"``.
+        align_corners (bool, optional): Changes the way the corners are treated
+            during ``"linear"`` and ``"cubic"`` upsampling.  See the note above and the
+            examples below for more details.  Default: ``False``.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+        >>> x = mx.arange(1, 5).reshape((1, 2, 2, 1))
+        >>> x
+        array([[[[1],
+                 [2]],
+                [[3],
+                 [4]]]], dtype=int32)
+        >>> n = nn.Upsample(scale_factor=2, mode='nearest')
+        >>> n(x).squeeze()
+        array([[1, 1, 2, 2],
+               [1, 1, 2, 2],
+               [3, 3, 4, 4],
+               [3, 3, 4, 4]], dtype=int32)
+        >>> b = nn.Upsample(scale_factor=2, mode='linear')
+        >>> b(x).squeeze()
+        array([[1, 1.25, 1.75, 2],
+               [1.5, 1.75, 2.25, 2.5],
+               [2.5, 2.75, 3.25, 3.5],
+               [3, 3.25, 3.75, 4]], dtype=float32)
+        >>> b = nn.Upsample(scale_factor=2, mode='linear', align_corners=True)
+        >>> b(x).squeeze()
+        array([[1, 1.33333, 1.66667, 2],
+               [1.66667, 2, 2.33333, 2.66667],
+               [2.33333, 2.66667, 3, 3.33333],
+               [3, 3.33333, 3.66667, 4]], dtype=float32)
+    """
+    def __init__(
+        self,
+        scale_factor: Union[float, Tuple],
+        mode: Literal["nearest", "linear", "cubic"] = ...,
+        align_corners: bool = ...,
+    ) -> None: ...
+    def __call__(self, x: mx.array) -> mx.array: ...

.mlx_typings/mlx/nn/losses.pyi

@@ -0,0 +1,419 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Literal, Optional
+
+import mlx.core as mx
+
+Reduction = Literal["none", "mean", "sum"]
+
+def cross_entropy(
+    logits: mx.array,
+    targets: mx.array,
+    weights: Optional[mx.array] = ...,
+    axis: int = ...,
+    label_smoothing: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    """
+    Computes the cross entropy loss.
+
+    Args:
+        logits (array): The unnormalized logits.
+        targets (array): The ground truth values. These can be class indices or
+            probabilities for each class. If the ``targets`` are class indices,
+            then ``targets`` shape should match the ``logits`` shape with
+            the ``axis`` dimension removed. If the ``targets`` are probabilities
+            (or one-hot encoded), then the ``targets`` shape should be the same as
+            the ``logits`` shape.
+        weights (array, optional): Optional weights for each target. Default: ``None``.
+        axis (int, optional): The axis over which to compute softmax. Default: ``-1``.
+        label_smoothing (float, optional): Label smoothing factor. Default: ``0``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+            ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed cross entropy loss.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+        >>>
+        >>> # Class indices as targets
+        >>> logits = mx.array([[2.0, -1.0], [-1.0, 2.0]])
+        >>> targets = mx.array([0, 1])
+        >>> nn.losses.cross_entropy(logits, targets)
+        array([0.0485873, 0.0485873], dtype=float32)
+        >>>
+        >>> # Probabilities (or one-hot vectors) as targets
+        >>> logits = mx.array([[2.0, -1.0], [-1.0, 2.0]])
+        >>> targets = mx.array([[0.9, 0.1], [0.1, 0.9]])
+        >>> nn.losses.cross_entropy(logits, targets)
+        array([0.348587, 0.348587], dtype=float32)
+    """
+
+def binary_cross_entropy(
+    inputs: mx.array,
+    targets: mx.array,
+    weights: Optional[mx.array] = ...,
+    with_logits: bool = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    """
+    Computes the binary cross entropy loss.
+
+    By default, this function takes the pre-sigmoid logits, which results in a faster
+    and more precise loss. For improved numerical stability when ``with_logits=False``,
+    the loss calculation clips the input probabilities (in log-space) to a minimum value
+    of ``-100``.
+
+    Args:
+        inputs (array): The predicted values. If ``with_logits`` is ``True``, then
+            ``inputs`` are unnormalized logits. Otherwise, ``inputs`` are probabilities.
+        targets (array): The binary target values in {0, 1}.
+        with_logits (bool, optional): Whether ``inputs`` are logits. Default: ``True``.
+        weights (array, optional): Optional weights for each target. Default: ``None``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'mean'``.
+
+    Returns:
+        array: The computed binary cross entropy loss.
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+
+        >>> logits = mx.array([0.105361, 0.223144, 1.20397, 0.916291])
+        >>> targets = mx.array([0, 0, 1, 1])
+        >>> loss = nn.losses.binary_cross_entropy(logits, targets, reduction="mean")
+        >>> loss
+        array(0.539245, dtype=float32)
+
+        >>> probs = mx.array([0.1, 0.1, 0.4, 0.4])
+        >>> targets = mx.array([0, 0, 1, 1])
+        >>> loss = nn.losses.binary_cross_entropy(probs, targets, with_logits=False, reduction="mean")
+        >>> loss
+        array(0.510826, dtype=float32)
+    """
+
+def l1_loss(
+    predictions: mx.array, targets: mx.array, reduction: Reduction = ...
+) -> mx.array:
+    """
+    Computes the L1 loss.
+
+    Args:
+        predictions (array): The predicted values.
+        targets (array): The target values.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'mean'``.
+
+    Returns:
+        array: The computed L1 loss.
+    """
+
+def mse_loss(
+    predictions: mx.array, targets: mx.array, reduction: Reduction = ...
+) -> mx.array:
+    """
+    Computes the mean squared error loss.
+
+    Args:
+        predictions (array): The predicted values.
+        targets (array): The target values.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'mean'``.
+
+    Returns:
+        array: The computed mean squared error loss.
+    """
+
+def nll_loss(
+    inputs: mx.array, targets: mx.array, axis: int = ..., reduction: Reduction = ...
+) -> mx.array:
+    """
+    Computes the negative log likelihood loss.
+
+    Args:
+        inputs (array): The predicted distribution in log space.
+        targets (array): The target values.
+        axis (int, optional): The distribution axis. Default: ``-1``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed NLL loss.
+    """
+
+def gaussian_nll_loss(
+    inputs: mx.array,
+    targets: mx.array,
+    vars: mx.array,
+    full: bool = ...,
+    eps: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    r"""
+    Computes the negative log likelihood loss for a Gaussian distribution.
+
+    The loss is given by:
+
+    .. math::
+        \frac{1}{2}\left(\log\left(\max\left(\text{vars},
+        \ \epsilon\right)\right) + \frac{\left(\text{inputs} - \text{targets} \right)^2}
+        {\max\left(\text{vars}, \ \epsilon \right)}\right) + \text{const.}
+
+    where ``inputs`` are the predicted means and ``vars`` are the the
+    predicted variances.
+
+    Args:
+        inputs (array): The predicted expectation of the Gaussian distribution.
+        targets (array): The target values (samples from the Gaussian distribution).
+        vars (array): The predicted variance of the Gaussian distribution.
+        full (bool, optional): Whether to include the constant term in the loss calculation.
+            Default: ``False``.
+        eps (float, optional): Small positive constant for numerical stability.
+            Default: ``1e-6``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The Gaussian NLL loss.
+    """
+
+def kl_div_loss(
+    inputs: mx.array, targets: mx.array, axis: int = ..., reduction: Reduction = ...
+) -> mx.array:
+    """
+    Computes the Kullback-Leibler divergence loss.
+
+    Computes the following when ``reduction == 'none'``:
+
+    .. code-block:: python
+
+        mx.exp(targets) * (targets - inputs).sum(axis)
+
+    Args:
+        inputs (array): Log probabilities for the predicted distribution.
+        targets (array): Log probabilities for the target distribution.
+        axis (int, optional): The distribution axis. Default: ``-1``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed Kullback-Leibler divergence loss.
+    """
+
+def smooth_l1_loss(
+    predictions: mx.array,
+    targets: mx.array,
+    beta: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    r"""
+    Computes the smooth L1 loss.
+
+    The smooth L1 loss is a variant of the L1 loss which replaces the absolute
+    difference with a squared difference when the absolute difference is less
+    than ``beta``.
+
+    The formula for the smooth L1 Loss is:
+
+    .. math::
+
+      l = \begin{cases}
+            0.5 (x - y)^2 / \beta, & \text{if } |x - y| < \beta \\
+            |x - y| - 0.5 \beta, & \text{otherwise}
+          \end{cases}
+
+    Args:
+        predictions (array): Predicted values.
+        targets (array): Ground truth values.
+        beta (float, optional): The threshold after which the loss changes
+          from the squared to the absolute difference. Default: ``1.0``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'mean'``.
+
+    Returns:
+        array: The computed smooth L1 loss.
+    """
+
+def triplet_loss(
+    anchors: mx.array,
+    positives: mx.array,
+    negatives: mx.array,
+    axis: int = ...,
+    p: int = ...,
+    margin: float = ...,
+    eps: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    r"""
+    Computes the triplet loss for a set of anchor, positive, and negative samples.
+    Margin is represented with alpha in the math section.
+
+    .. math::
+
+       \max\left(\|A - P\|_p - \|A - N\|_p + \alpha, 0\right)
+
+    Args:
+        anchors (array): The anchor samples.
+        positives (array): The positive samples.
+        negatives (array): The negative samples.
+        axis (int, optional): The distribution axis. Default: ``-1``.
+        p (int, optional): The norm degree for pairwise distance. Default: ``2``.
+        margin (float, optional): Margin for the triplet loss. Defaults to ``1.0``.
+        eps (float, optional): Small positive constant to prevent numerical instability. Defaults to ``1e-6``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: Computed triplet loss. If reduction is "none", returns a tensor of the same shape as input;
+                  if reduction is "mean" or "sum", returns a scalar tensor.
+    """
+
+def hinge_loss(
+    inputs: mx.array, targets: mx.array, reduction: Reduction = ...
+) -> mx.array:
+    r"""
+    Computes the hinge loss between inputs and targets.
+
+    .. math::
+
+       \text{hinge}(y, y_{\text{pred}}) = \max(0, 1 - y \cdot y_{\text{pred}})
+
+
+    Args:
+        inputs (array): The predicted values.
+        targets (array): The target values. They should be -1 or 1.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed hinge loss.
+    """
+
+def huber_loss(
+    inputs: mx.array, targets: mx.array, delta: float = ..., reduction: Reduction = ...
+) -> mx.array:
+    r"""
+    Computes the Huber loss between inputs and targets.
+
+    .. math::
+
+        l_{\delta}(a) =
+        \left\{ \begin{array}{ll}
+            \frac{1}{2} a^2 & \text{for } |a| \leq \delta, \\
+            \delta \left( |a| - \frac{1}{2} \delta \right) & \text{otherwise.}
+        \end{array} \right.
+
+    Args:
+        inputs (array): The predicted values.
+        targets (array): The target values.
+        delta (float, optional): The threshold at which to change between L1 and L2 loss.
+          Default: ``1.0``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed Huber loss.
+    """
+
+def log_cosh_loss(
+    inputs: mx.array, targets: mx.array, reduction: Reduction = ...
+) -> mx.array:
+    r"""
+    Computes the log cosh loss between inputs and targets.
+
+    Logcosh acts like L2 loss for small errors, ensuring stable gradients,
+    and like the L1 loss for large errors, reducing sensitivity to outliers. This
+    dual behavior offers a balanced, robust approach for regression tasks.
+
+    .. math::
+
+       \text{logcosh}(y_{\text{true}}, y_{\text{pred}}) =
+            \frac{1}{n} \sum_{i=1}^{n}
+            \log(\cosh(y_{\text{pred}}^{(i)} - y_{\text{true}}^{(i)}))
+
+
+    Args:
+        inputs (array): The predicted values.
+        targets (array): The target values.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed log cosh loss.
+    """
+
+def cosine_similarity_loss(
+    x1: mx.array,
+    x2: mx.array,
+    axis: int = ...,
+    eps: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    r"""
+    Computes the cosine similarity between the two inputs.
+
+    The cosine similarity loss is given by
+
+    .. math::
+
+        \frac{x_1 \cdot x_2}{\max(\|x_1\|  \cdot \|x_2\|, \epsilon)}
+
+    Args:
+        x1 (mx.array): The first set of inputs.
+        x2 (mx.array): The second set of inputs.
+        axis (int, optional): The embedding axis. Default: ``1``.
+        eps (float, optional): The minimum value of the denominator used for
+          numerical stability. Default: ``1e-8``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+          ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        mx.array: The computed cosine similarity loss.
+    """
+
+def margin_ranking_loss(
+    inputs1: mx.array,
+    inputs2: mx.array,
+    targets: mx.array,
+    margin: float = ...,
+    reduction: Reduction = ...,
+) -> mx.array:
+    r"""
+    Calculate the margin ranking loss that loss given inputs :math:`x_1`, :math:`x_2` and a label
+    :math:`y` (containing 1 or -1).
+
+    The loss is given by:
+
+    .. math::
+        \text{loss} = \max (0, -y * (x_1 - x_2) + \text{margin})
+
+    Where :math:`y` represents ``targets``, :math:`x_1` represents ``inputs1`` and :math:`x_2`
+    represents ``inputs2``.
+
+    Args:
+        inputs1 (array): Scores for the first input.
+        inputs2 (array): Scores for the second input.
+        targets (array): Labels indicating whether samples in ``inputs1`` should be ranked higher
+            than samples in ``inputs2``. Values should be 1 or -1.
+        margin (float, optional): The margin by which the scores should be separated.
+            Default: ``0.0``.
+        reduction (str, optional): Specifies the reduction to apply to the output:
+            ``'none'`` | ``'mean'`` | ``'sum'``. Default: ``'none'``.
+
+    Returns:
+        array: The computed margin ranking loss.
+
+    Examples:
+        >>> import mlx.core as mx
+        >>> import mlx.nn as nn
+        >>> targets = mx.array([1, 1, -1])
+        >>> inputs1 = mx.array([-0.573409, -0.765166, -0.0638])
+        >>> inputs2 = mx.array([0.75596, 0.225763, 0.256995])
+        >>> loss = nn.losses.margin_ranking_loss(inputs1, inputs2, targets)
+        >>> loss
+        array(0.773433, dtype=float32)
+    """

.mlx_typings/mlx/nn/utils.pyi

@@ -0,0 +1,73 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any, Callable, Optional
+
+import mlx.core as mx
+
+from .layers.base import Module
+
+def value_and_grad(
+    model: Module, fn: Callable
+):  # -> _Wrapped[..., Any, ..., tuple[Any, Any]]:
+    """Transform the passed function ``fn`` to a function that computes the
+    gradients of ``fn`` wrt the model's trainable parameters and also its
+    value.
+
+    Args:
+        model (Module): The model whose trainable parameters to compute
+                               gradients for
+        fn (Callable): The scalar function to compute gradients for
+
+    Returns:
+        A callable that returns the value of ``fn`` and the gradients wrt the
+        trainable parameters of ``model``
+    """
+
+def checkpoint(
+    module: Module, fn: Optional[Callable] = ...
+):  # -> _Wrapped[..., Any, ..., Any]:
+    """Transform the passed callable to one that performs gradient
+    checkpointing with respect to the trainable parameters of the module (and
+    the callable's inputs).
+
+    Args:
+        module (Module): The module for whose parameters we will be
+            performing gradient checkpointing.
+        fn (Callable, optional): The function to checkpoint. If not provided it
+            defaults to the provided module.
+
+    Returns:
+        A callable that saves the inputs and outputs during the forward pass
+        and recomputes all intermediate states during the backward pass.
+    """
+
+def average_gradients(
+    gradients: Any,
+    group: Optional[mx.distributed.Group] = ...,
+    all_reduce_size: int = ...,
+    communication_type: Optional[mx.Dtype] = ...,
+    communication_stream: Optional[mx.Stream] = ...,
+):  # -> Any:
+    """Average the gradients across the distributed processes in the passed group.
+
+    This helper enables concatenating several gradients of small arrays to one
+    big all reduce call for better networking performance.
+
+    Args:
+        gradients (Any): The Python tree containing the gradients (it should
+            have the same structure across processes)
+        group (Optional[mlx.core.distributed.Group]): The group of processes to
+            average the gradients. If set to ``None`` the global group is used.
+            Default: ``None``.
+        all_reduce_size (int): Group arrays until their size in bytes exceeds
+            this number. Perform one communication step per group of arrays. If
+            less or equal to 0 array grouping is disabled. Default: ``32MiB``.
+        communication_type (Optional[mlx.core.Dtype]): If provided cast to this
+            type before performing the communication. Typically cast to a
+            smaller float to reduce the communication size. Default: ``None``.
+        communication_stream (Optional[mlx.core.Stream]): The stream to usse
+            for the communication. If unspecified the default communication
+            stream is used which can vary by back-end. Default: ``None``.
+    """

.mlx_typings/mlx/utils.pyi

@@ -0,0 +1,189 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+from mlx.core import MX_ARRAY_TREE
+
+def tree_map(
+    fn: Callable, tree: Any, *rest: Any, is_leaf: Optional[Callable] = ...
+) -> Any:
+    """Applies ``fn`` to the leaves of the Python tree ``tree`` and
+    returns a new collection with the results.
+
+    If ``rest`` is provided, every item is assumed to be a superset of ``tree``
+    and the corresponding leaves are provided as extra positional arguments to
+    ``fn``. In that respect, :meth:`tree_map` is closer to :func:`itertools.starmap`
+    than to :func:`map`.
+
+    The keyword argument ``is_leaf`` decides what constitutes a leaf from
+    ``tree`` similar to :func:`tree_flatten`.
+
+    .. code-block:: python
+
+        import mlx.nn as nn
+        from mlx.utils import tree_map
+
+        model = nn.Linear(10, 10)
+        print(model.parameters().keys())
+        # dict_keys(['weight', 'bias'])
+
+        # square the parameters
+        model.update(tree_map(lambda x: x*x, model.parameters()))
+
+    Args:
+        fn (callable): The function that processes the leaves of the tree.
+        tree (Any): The main Python tree that will be iterated upon.
+        rest (tuple[Any]): Extra trees to be iterated together with ``tree``.
+        is_leaf (callable, optional): An optional callable that returns ``True``
+           if the passed object is considered a leaf or ``False`` otherwise.
+
+    Returns:
+        A Python tree with the new values returned by ``fn``.
+    """
+
+def tree_map_with_path(
+    fn: Callable,
+    tree: Any,
+    *rest: Any,
+    is_leaf: Optional[Callable] = ...,
+    path: Optional[Any] = ...,
+) -> Any:
+    """Applies ``fn`` to the path and leaves of the Python tree ``tree`` and
+    returns a new collection with the results.
+
+    This function is the same :func:`tree_map` but the ``fn`` takes the path as
+    the first argument followed by the remaining tree nodes.
+
+    Args:
+        fn (callable): The function that processes the leaves of the tree.
+        tree (Any): The main Python tree that will be iterated upon.
+        rest (tuple[Any]): Extra trees to be iterated together with ``tree``.
+        is_leaf (Optional[Callable]): An optional callable that returns ``True``
+           if the passed object is considered a leaf or ``False`` otherwise.
+        path (Optional[Any]): Prefix will be added to the result.
+
+    Returns:
+        A Python tree with the new values returned by ``fn``.
+
+    Example:
+        >>> from mlx.utils import tree_map_with_path
+        >>> tree = {"model": [{"w": 0, "b": 1}, {"w": 0, "b": 1}]}
+        >>> new_tree = tree_map_with_path(lambda path, _: print(path), tree)
+        model.0.w
+        model.0.b
+        model.1.w
+        model.1.b
+    """
+
+def tree_flatten(
+    tree: Any,
+    prefix: str = ...,
+    is_leaf: Optional[Callable] = ...,
+    destination: Optional[Union[List[Tuple[str, Any]], Dict[str, Any]]] = ...,
+) -> Union[List[Tuple[str, Any]], Dict[str, Any]]:
+    """Flattens a Python tree to a list of key, value tuples.
+
+    The keys are using the dot notation to define trees of arbitrary depth and
+    complexity.
+
+    .. code-block:: python
+
+        from mlx.utils import tree_flatten
+
+        print(tree_flatten([[[0]]]))
+        # [("0.0.0", 0)]
+
+        print(tree_flatten([[[0]]], prefix=".hello"))
+        # [("hello.0.0.0", 0)]
+
+        tree_flatten({"a": {"b": 1}}, destination={})
+        {"a.b": 1}
+
+    .. note::
+       Dictionaries should have keys that are valid Python identifiers.
+
+    Args:
+        tree (Any): The Python tree to be flattened.
+        prefix (str): A prefix to use for the keys. The first character is
+            always discarded.
+        is_leaf (callable): An optional callable that returns True if the
+            passed object is considered a leaf or False otherwise.
+        destination (list or dict, optional): A list or dictionary to store the
+            flattened tree. If None an empty list will be used. Default: ``None``.
+
+    Returns:
+        Union[List[Tuple[str, Any]], Dict[str, Any]]: The flat representation of
+            the Python tree.
+    """
+
+def tree_unflatten(tree: Union[List[Tuple[str, Any]], Dict[str, Any]]) -> Any:
+    """Recreate a Python tree from its flat representation.
+
+    .. code-block:: python
+
+        from mlx.utils import tree_unflatten
+
+        d = tree_unflatten([("hello.world", 42)])
+        print(d)
+        # {"hello": {"world": 42}}
+
+        d = tree_unflatten({"hello.world": 42})
+        print(d)
+        # {"hello": {"world": 42}}
+
+    Args:
+        tree (list[tuple[str, Any]] or dict[str, Any]): The flat representation of a Python tree.
+           For instance as returned by :meth:`tree_flatten`.
+
+    Returns:
+        A Python tree.
+    """
+
+def tree_reduce(
+    fn: Callable[[Any, Any], Any],
+    tree: list[MX_ARRAY_TREE] | tuple[MX_ARRAY_TREE, ...] | dict[str, MX_ARRAY_TREE],
+    initializer=...,
+    is_leaf=...,
+) -> None:
+    """Applies a reduction to the leaves of a Python tree.
+
+    This function reduces Python trees into an accumulated result by applying
+    the provided function ``fn`` to the leaves of the tree.
+
+    Example:
+        >>> from mlx.utils import tree_reduce
+        >>> tree = {"a": [1, 2, 3], "b": [4, 5]}
+        >>> tree_reduce(lambda acc, x: acc + x, tree, 0)
+        15
+
+    Args:
+        fn (callable): The reducer function that takes two arguments (accumulator,
+            current value) and returns the updated accumulator.
+        tree (Any): The Python tree to reduce. It can be any nested combination of
+            lists, tuples, or dictionaries.
+        initializer (Any, optional): The initial value to start the reduction. If
+            not provided, the first leaf value is used.
+        is_leaf (callable, optional): A function to determine if an object is a
+            leaf, returning ``True`` for leaf nodes and ``False`` otherwise.
+
+    Returns:
+        Any: The accumulated value.
+    """
+
+def tree_merge(
+    tree_a, tree_b, merge_fn=...
+):  # -> dict[Any, Any] | list[Any] | tuple[Any, *tuple[Any, ...]] | tuple[Any, ...]:
+    """Merge two Python trees in one containing the values of both. It can be
+    thought of as a deep dict.update method.
+
+    Args:
+        tree_a (Any): The first Python tree.
+        tree_b (Any): The second Python tree.
+        merge_fn (callable, optional): A function to merge leaves.
+
+    Returns:
+        The Python tree containing the values of both ``tree_a`` and
+        ``tree_b``.
+    """

.mlx_typings/mlx_lm/__init__.pyi

@@ -0,0 +1,3 @@
+import models as models
+import tokenizer_utils as tokenizer_utils
+from generate import *

.mlx_typings/mlx_lm/convert.pyi

@@ -0,0 +1,45 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import argparse
+from typing import Callable, Optional, Union
+
+import mlx.nn as nn
+
+def mixed_quant_predicate_builder(
+    recipe: str, model: nn.Module, group_size: int = ...
+) -> Callable[[str, nn.Module, dict], Union[bool, dict]]: ...
+
+QUANT_RECIPES = ...
+MODEL_CONVERSION_DTYPES = ...
+
+def convert(
+    hf_path: str,
+    mlx_path: str = ...,
+    quantize: bool = ...,
+    q_group_size: int = ...,
+    q_bits: int = ...,
+    q_mode: str = ...,
+    dtype: Optional[str] = ...,
+    upload_repo: str = ...,
+    revision: Optional[str] = ...,
+    dequantize: bool = ...,
+    quant_predicate: Optional[
+        Union[Callable[[str, nn.Module, dict], Union[bool, dict]], str]
+    ] = ...,
+    trust_remote_code: bool = ...,
+):  # -> None:
+    ...
+def configure_parser() -> argparse.ArgumentParser:
+    """
+    Configures and returns the argument parser for the script.
+
+    Returns:
+        argparse.ArgumentParser: Configured argument parser.
+    """
+
+def main():  # -> None:
+    ...
+
+if __name__ == "__main__": ...

.mlx_typings/mlx_lm/generate.pyi

@@ -0,0 +1,324 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import contextlib
+from dataclasses import dataclass
+from typing import Any, Callable, Generator, List, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+from transformers import PreTrainedTokenizer
+
+from .tokenizer_utils import TokenizerWrapper
+
+DEFAULT_PROMPT = ...
+DEFAULT_MAX_TOKENS = ...
+DEFAULT_TEMP = ...
+DEFAULT_TOP_P = ...
+DEFAULT_MIN_P = ...
+DEFAULT_TOP_K = ...
+DEFAULT_XTC_PROBABILITY = ...
+DEFAULT_XTC_THRESHOLD = ...
+DEFAULT_MIN_TOKENS_TO_KEEP = ...
+DEFAULT_SEED = ...
+DEFAULT_MODEL = ...
+DEFAULT_QUANTIZED_KV_START = ...
+
+def str2bool(string):  # -> bool:
+    ...
+def setup_arg_parser():  # -> ArgumentParser:
+    """Set up and return the argument parser."""
+
+generation_stream = ...
+
+@contextlib.contextmanager
+def wired_limit(
+    model: nn.Module, streams: Optional[List[mx.Stream]] = ...
+):  # -> Generator[None, Any, None]:
+    """
+    A context manager to temporarily change the wired limit.
+
+    Note, the wired limit should not be changed during an async eval.  If an
+    async eval could be running pass in the streams to synchronize with prior
+    to exiting the context manager.
+    """
+@dataclass
+class GenerationResponse:
+    """
+    The output of :func:`stream_generate`.
+
+    Args:
+        text (str): The next segment of decoded text. This can be an empty string.
+        token (int): The next token.
+        from_draft (bool): Whether the token was generated by the draft model.
+        logprobs (mx.array): A vector of log probabilities.
+        prompt_tokens (int): The number of tokens in the prompt.
+        prompt_tps (float): The prompt processing tokens-per-second.
+        generation_tokens (int): The number of generated tokens.
+        generation_tps (float): The tokens-per-second for generation.
+        peak_memory (float): The peak memory used so far in GB.
+        finish_reason (str): The reason the response is being sent: "length", "stop" or `None`
+    """
+
+    text: str
+    token: int
+    logprobs: mx.array
+    from_draft: bool
+    prompt_tokens: int
+    prompt_tps: float
+    generation_tokens: int
+    generation_tps: float
+    peak_memory: float
+    finish_reason: Optional[str] = ...
+
+def maybe_quantize_kv_cache(
+    prompt_cache, quantized_kv_start, kv_group_size, kv_bits
+):  # -> None:
+    ...
+def generate_step(
+    prompt: mx.array,
+    model: nn.Module,
+    *,
+    max_tokens: int = ...,
+    sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+    logits_processors: Optional[List[Callable[[mx.array, mx.array], mx.array]]] = ...,
+    max_kv_size: Optional[int] = ...,
+    prompt_cache: Optional[Any] = ...,
+    prefill_step_size: int = ...,
+    kv_bits: Optional[int] = ...,
+    kv_group_size: int = ...,
+    quantized_kv_start: int = ...,
+    prompt_progress_callback: Optional[Callable[[int], int]] = ...,
+    input_embeddings: Optional[mx.array] = ...,
+) -> Generator[Tuple[mx.array, mx.array], None, None]:
+    """
+    A generator producing token ids based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        max_tokens (int): The maximum number of tokens. Use``-1`` for an infinite
+          generator. Default: ``256``.
+        sampler (Callable[mx.array, mx.array], optional): A sampler for sampling a
+          token from a vector of log probabilities. Default: ``None``.
+        logits_processors (List[Callable[[mx.array, mx.array], mx.array]], optional):
+          A list of functions that take tokens and logits and return the processed
+          logits. Default: ``None``.
+        max_kv_size (int, optional): Maximum size of the key-value cache. Old
+          entries (except the first 4 tokens) will be overwritten.
+        prompt_cache (List[Any], optional): A pre-computed prompt cache. Note, if
+          provided, the cache will be updated in place.
+        prefill_step_size (int): Step size for processing the prompt.
+        kv_bits (int, optional): Number of bits to use for KV cache quantization.
+          None implies no cache quantization. Default: ``None``.
+        kv_group_size (int): Group size for KV cache quantization. Default: ``64``.
+        quantized_kv_start (int): Step to begin using a quantized KV cache.
+           when ``kv_bits`` is non-None. Default: ``0``.
+        prompt_progress_callback (Callable[[int], int]): A call-back which takes the
+           prompt tokens processed so far and the total number of prompt tokens.
+        input_embeddings (mx.array, optional): Input embeddings to use instead of or in
+          conjunction with prompt tokens. Default: ``None``.
+
+    Yields:
+        Tuple[mx.array, mx.array]: One token and a vector of log probabilities.
+    """
+
+def speculative_generate_step(
+    prompt: mx.array,
+    model: nn.Module,
+    draft_model: nn.Module,
+    *,
+    num_draft_tokens: int = ...,
+    max_tokens: int = ...,
+    sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+    logits_processors: Optional[List[Callable[[mx.array, mx.array], mx.array]]] = ...,
+    prompt_cache: Optional[Any] = ...,
+    prefill_step_size: int = ...,
+    kv_bits: Optional[int] = ...,
+    kv_group_size: int = ...,
+    quantized_kv_start: int = ...,
+) -> Generator[Tuple[mx.array, mx.array, bool], None, None]:
+    """
+    A generator producing token ids based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        draft_model (nn.Module): The draft model for speculative decoding.
+        num_draft_tokens (int, optional): The number of draft tokens for
+          speculative decoding. Default: ``2``.
+        max_tokens (int): The maximum number of tokens. Use``-1`` for an infinite
+          generator. Default: ``256``.
+        sampler (Callable[[mx.array], mx.array], optional): A sampler for sampling a
+          token from a vector of log probabilities. Default: ``None``.
+        logits_processors (List[Callable[[mx.array, mx.array], mx.array]], optional):
+          A list of functions that take tokens and logits and return the processed
+          logits. Default: ``None``.
+        prompt_cache (List[Any], optional): A pre-computed prompt cache. Note, if
+          provided, the cache will be updated in place. The cache must be trimmable.
+        prefill_step_size (int): Step size for processing the prompt.
+        kv_bits (int, optional): Number of bits to use for KV cache quantization.
+          None implies no cache quantization. Default: ``None``.
+        kv_group_size (int): Group size for KV cache quantization. Default: ``64``.
+        quantized_kv_start (int): Step to begin using a quantized KV cache.
+           when ``kv_bits`` is non-None. Default: ``0``.
+
+    Yields:
+        Tuple[mx.array, mx.array, bool]: One token, a vector of log probabilities,
+          and a bool indicating if the token was generated by the draft model
+    """
+
+def stream_generate(
+    model: nn.Module,
+    tokenizer: Union[PreTrainedTokenizer, TokenizerWrapper],
+    prompt: Union[str, mx.array, List[int]],
+    max_tokens: int = ...,
+    draft_model: Optional[nn.Module] = ...,
+    **kwargs: object,
+) -> Generator[GenerationResponse, None, None]:
+    """
+    A generator producing text based on the given prompt from the model.
+
+    Args:
+        model (nn.Module): The model to use for generation.
+        tokenizer (PreTrainedTokenizer): The tokenizer.
+        prompt (Union[str, mx.array, List[int]]): The input prompt string or
+          integer tokens.
+        max_tokens (int): The maximum number of tokens to generate.
+          Default: ``256``.
+        draft_model (Optional[nn.Module]): An optional draft model. If provided
+          then speculative decoding is used. The draft model must use the same
+          tokenizer as the main model. Default: ``None``.
+        kwargs: The remaining options get passed to :func:`generate_step`.
+          See :func:`generate_step` for more details.
+
+    Yields:
+        GenerationResponse: An instance containing the generated text segment and
+            associated metadata. See :class:`GenerationResponse` for details.
+    """
+
+def generate(
+    model: nn.Module,
+    tokenizer: Union[PreTrainedTokenizer, TokenizerWrapper],
+    prompt: Union[str, List[int]],
+    verbose: bool = ...,
+    **kwargs,
+) -> str:
+    """
+    Generate a complete response from the model.
+
+    Args:
+       model (nn.Module): The language model.
+       tokenizer (PreTrainedTokenizer): The tokenizer.
+       prompt (Union[str, List[int]]): The input prompt string or integer tokens.
+       verbose (bool): If ``True``, print tokens and timing information.
+           Default: ``False``.
+       kwargs: The remaining options get passed to :func:`stream_generate`.
+          See :func:`stream_generate` for more details.
+    """
+@dataclass
+class BatchStats:
+    """
+    An data object to hold generation stats.
+
+    Args:
+        prompt_tokens (int): The number of prompt tokens processed.
+        prompt_tps (float): The prompt processing tokens-per-second.
+        prompt_time (float): The time in seconds spent in prompt processing.
+        generation_tokens (int): The number of generated tokens.
+        generation_tps (float): The tokens-per-second for generation.
+        generation_time (float): The time in seconds spent in generation .
+        peak_memory (float): The peak memory used so far in GB.
+    """
+
+    prompt_tokens: int = ...
+    prompt_tps: float = ...
+    prompt_time: float = ...
+    generation_tokens: int = ...
+    generation_tps: float = ...
+    generation_time: float = ...
+    peak_memory: float = ...
+
+@dataclass
+class BatchResponse:
+    """
+    An data object to hold a batch generation response.
+
+    Args:
+        texts: (List[str]): The generated text for each prompt.
+        stats (BatchStats): Statistics about the generation.
+    """
+
+    texts: List[str]
+    stats: BatchStats
+
+@dataclass
+class Batch:
+    uids: List[int]
+    y: mx.array
+    logprobs: mx.array
+    max_tokens: List[int]
+    num_tokens: List[int]
+    cache: List[Any]
+    def __len__(self):  # -> int:
+        ...
+    def filter(self, keep_idx: List[int]):  # -> None:
+        ...
+    def extend(self, other):  # -> None:
+        ...
+
+class BatchGenerator:
+    @dataclass
+    class Response:
+        uid: int
+        token: int
+        logprobs: mx.array
+        finish_reason: Optional[str]
+
+    def __init__(
+        self,
+        model,
+        max_tokens: int = ...,
+        stop_tokens: Optional[set] = ...,
+        sampler: Optional[Callable[[mx.array], mx.array]] = ...,
+        completion_batch_size: int = ...,
+        prefill_batch_size: int = ...,
+        prefill_step_size: int = ...,
+    ) -> None: ...
+    def insert(
+        self, prompts, max_tokens: Union[List[int], int, None] = ...
+    ):  # -> list[Any]:
+        ...
+    def stats(self):  # -> BatchStats:
+        ...
+    def next(self):  # -> list[Any]:
+        ...
+
+def batch_generate(
+    model,
+    tokenizer,
+    prompts: List[int],
+    max_tokens: Union[int, List[int]] = ...,
+    verbose: bool = ...,
+    **kwargs,
+) -> BatchResponse:
+    """
+    Generate responses for the given batch of prompts.
+
+    Args:
+       model (nn.Module): The language model.
+       tokenizer (PreTrainedTokenizer): The tokenizer.
+       prompt (List[List[int]]): The input prompts.
+       verbose (bool): If ``True``, print tokens and timing information.
+          Default: ``False``.
+       max_tokens (Union[int, List[int]): Maximum number of output tokens. This
+          can be per prompt if a list is provided.
+       kwargs: The remaining options get passed to :obj:`BatchGenerator`.
+          See :obj:`BatchGenerator` for more details.
+    """
+
+def main():  # -> None:
+    ...
+
+if __name__ == "__main__": ...

.mlx_typings/mlx_lm/models/__init__.pyi

@@ -0,0 +1 @@
+import cache as cache

.mlx_typings/mlx_lm/models/base.pyi

@@ -0,0 +1,47 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+
+@dataclass
+class BaseModelArgs:
+    @classmethod
+    def from_dict(cls, params):  # -> Self:
+        ...
+
+def create_causal_mask(
+    N: int,
+    offset: int = ...,
+    window_size: Optional[int] = ...,
+    right_padding: Optional[mx.array] = ...,
+    left_padding: Optional[mx.array] = ...,
+):  # -> array:
+    ...
+def create_attention_mask(
+    h, cache=..., window_size: Optional[int] = ..., return_array: bool = ...
+):  # -> array | Literal['causal'] | None:
+    ...
+def create_ssm_mask(h, cache=...):  # -> None:
+    ...
+def quantized_scaled_dot_product_attention(
+    queries: mx.array,
+    q_keys: tuple[mx.array, mx.array, mx.array],
+    q_values: tuple[mx.array, mx.array, mx.array],
+    scale: float,
+    mask: Optional[mx.array],
+    group_size: int = ...,
+    bits: int = ...,
+) -> mx.array: ...
+def scaled_dot_product_attention(
+    queries,
+    keys,
+    values,
+    cache,
+    scale: float,
+    mask: Optional[mx.array],
+    sinks: Optional[mx.array] = ...,
+) -> mx.array: ...

.mlx_typings/mlx_lm/models/bitlinear_layers.pyi

@@ -0,0 +1,26 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import mlx.nn as nn
+
+def bitnet_quantize(model, quantization_config: dict): ...
+def make_bitlinear_kernel():
+    """
+    Custom Metal kernel that performs matrix multiplication directly on
+    packed weights and scales the output. This eliminates the need to
+    store unpacked weights in memory.
+    """
+
+_bitlinear_kernel = ...
+
+class BitLinear(nn.Module):
+    """
+    BitLinear module with memory-efficient weight handling.
+    """
+    def __init__(
+        self, in_features, out_features, bias=..., invert_weight_scales=...
+    ) -> None: ...
+    def execute_matmul_kernel(self, x, packed_weights): ...
+    def __call__(self, x):  # -> array:
+        ...

.mlx_typings/mlx_lm/models/cache.pyi

@@ -0,0 +1,357 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from typing import Any, Dict, List, Optional, Protocol, Literal, Self
+
+import mlx.nn as nn
+from mlx.core import array
+import mlx.core as mx
+
+class Cache(Protocol):
+    keys: mx.array
+    values: mx.array
+    def update_and_fetch(self, keys: mx.array, values: mx.array) -> None: ...
+    @property
+    def state(self) -> tuple[mx.array, mx.array]: ...
+    @state.setter
+    def state(self, v) -> None: ...
+
+def make_prompt_cache(
+    model: nn.Module, max_kv_size: Optional[int] = ...
+) -> List[Cache | Any]:
+    """
+    Construct the model's cache for use in generation.
+
+    This function will defer the cache construction to the model if it has a
+    ``make_cache`` method, otherwise it will make a default KV cache.
+
+    Args:
+        model (nn.Module): The language model.
+        max_kv_size (Optional[int]): If provided and the model does not have a
+            ``make_cache`` method, a ``RotatingKVCache`` is used with a maximum
+            size of ``max_kv_size``
+    """
+
+def save_prompt_cache(
+    file_name: str, cache: List[Cache], metadata: Dict[str, str] = ...
+) -> None:
+    """
+    Save a pre-computed prompt cache to a file.
+
+    Args:
+        file_name (str): The ``.safetensors`` file name.
+        cache (List[Any]): The model state.
+        metadata (Dict[str, str]): Optional metadata to save along with model
+            state.
+    """
+
+def load_prompt_cache(file_name: str, return_metadata=...) -> array:
+    """
+    Load a prompt cache from a file.
+
+    Args:
+        file_name (str): The ``.safetensors`` file name.
+        return_metadata (bool): Whether or not to return metadata.
+            Default: ``False``.
+
+    Returns:
+        List[Any] or Tuple[List[Any], Dict[str, str]]: The prompt cache and
+            the metadata if requested.
+    """
+
+def can_trim_prompt_cache(cache: List[Cache]) -> bool:
+    """
+    Check if model's cache can be trimmed.
+    """
+
+def trim_prompt_cache(cache: List[Cache], num_tokens: int) -> List[Cache]:
+    """
+    Trim the model's cache by the given number of tokens.
+
+    This function will trim the cache if possible (in-place) and return the
+    number of tokens that were trimmed.
+
+    Args:
+        cache (List[Any]): The model's cache.
+        num_tokens (int): The number of tokens to trim.
+
+    Returns:
+        (int): The number of tokens that were trimmed.
+    """
+
+def create_attention_mask(
+    N: int, offset: int, return_array: bool, window_size: Optional[int]
+) -> array | Literal["causal"] | None: ...
+
+class _BaseCache(Cache):
+    keys: mx.array
+    values: mx.array
+    @property
+    def state(self) -> tuple[mx.array, mx.array]: ...
+    @state.setter
+    def state(self, v) -> None: ...
+    @property
+    def meta_state(self) -> Literal[""]: ...
+    @meta_state.setter
+    def meta_state(self, v) -> None: ...
+    def is_trimmable(self) -> Literal[False]: ...
+    @classmethod
+    def from_state(cls, state, meta_state) -> Self: ...
+
+class ConcatenateKVCache(_BaseCache):
+    """ConcatenateKVCache the simplest KV cache implementation.
+
+    Can be used as a mock KV cache or when large blocks are being processed at
+    a time in which case KVCache isn't necessarily faster. Consider using the
+    KVCache with a larger step size before using this cache.
+    """
+    def __init__(self) -> None: ...
+    def update_and_fetch(self, keys, values):  # -> tuple[Any | array, Any | array]:
+        ...
+    @property
+    def state(self):  # -> tuple[Any | array | None, Any | array | None]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    def is_trimmable(self):  # -> Literal[True]:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
+        ...
+
+class QuantizedKVCache(_BaseCache):
+    step = ...
+    def __init__(self, group_size: int = ..., bits: int = ...) -> None: ...
+    def update_and_fetch(self, keys, values):  # -> Any:
+        ...
+    @property
+    def state(
+        self,
+    ):  # -> tuple[Any | tuple[array, array, array] | None, Any | tuple[array, array, array] | None] | Any:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    @property
+    def meta_state(self):  # -> tuple[str, ...]:
+        ...
+    @meta_state.setter
+    def meta_state(self, v):  # -> None:
+        ...
+    def is_trimmable(self):  # -> Literal[True]:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
+        ...
+
+class KVCache(_BaseCache):
+    step = ...
+    def __init__(self) -> None: ...
+    def update_and_fetch(self, keys, values):  # -> tuple[array | Any, array | Any]:
+        ...
+    @property
+    def state(
+        self,
+    ) -> tuple[array, array]: ...
+    @state.setter
+    def state(self, v) -> None: ...
+    def is_trimmable(self):  # -> Literal[True]:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ...
+    ) -> QuantizedKVCache: ...
+    def make_mask(self, *args, **kwargs):  # -> array | Literal['causal'] | None:
+        ...
+
+class RotatingKVCache(_BaseCache):
+    step = ...
+    def __init__(self, max_size, keep=...) -> None: ...
+    def update_and_fetch(
+        self, keys, values
+    ):  # -> tuple[array | Any, array | Any] | tuple[array | Any, array | Any | None]:
+        ...
+    @property
+    def state(
+        self,
+    ):  # -> tuple[Any | array, Any | array] | tuple[Any | array | None, Any | array | None]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    @property
+    def meta_state(self):  # -> tuple[str, ...]:
+        ...
+    @meta_state.setter
+    def meta_state(self, v):  # -> None:
+        ...
+    def is_trimmable(self):  # -> bool:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ...
+    ) -> QuantizedKVCache: ...
+    def make_mask(
+        self, N: int, window_size: Optional[int] = ..., return_array: bool = ...
+    ):  # -> array | Literal['causal'] | None:
+        ...
+
+class ArraysCache(_BaseCache):
+    def __init__(self, size, left_padding: Optional[List[int]] = ...) -> None: ...
+    def __setitem__(self, idx, value):  # -> None:
+        ...
+    def __getitem__(self, idx): ...
+    @property
+    def state(self):  # -> list[Any | array] | list[array]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    def filter(self, batch_indices):  # -> None:
+        """
+        In-place filter to keep just the given indices in the cache.
+        """
+
+    def extend(self, other):  # -> None:
+        """
+        In-place extend this cache with the other cache.
+        """
+
+    def make_mask(self, N: int):  # -> array | None:
+        ...
+
+class MambaCache(ArraysCache):
+    def __init__(self, left_padding: Optional[List[int]] = ...) -> None: ...
+
+class ChunkedKVCache(KVCache):
+    def __init__(self, chunk_size) -> None: ...
+    def maybe_trim_front(self):  # -> None:
+        ...
+    def update_and_fetch(self, keys, values):  # -> tuple[array, array]:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    @property
+    def meta_state(self):  # -> tuple[str, ...]:
+        ...
+    @meta_state.setter
+    def meta_state(self, v):  # -> None:
+        ...
+
+class CacheList(_BaseCache):
+    def __init__(self, *caches) -> None: ...
+    def __getitem__(self, idx): ...
+    def is_trimmable(self):  # -> bool:
+        ...
+    def trim(self, n): ...
+    @property
+    def state(self):  # -> list[Any]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    def filter(self, batch_indices):  # -> None:
+        """
+        In-place filter to keep just the given indices in the cache.
+        """
+
+    def extend(self, other):  # -> None:
+        """
+        In-place extend this cache with the other cache.
+        """
+
+class BatchKVCache(_BaseCache):
+    step = ...
+    def __init__(self, left_padding: List[int]) -> None:
+        """
+        The BatchKV cache expects inputs to be left-padded.
+
+        E.g. the following prompts:
+
+            [1, 3, 5]
+            [7]
+            [2, 6, 8, 9]
+
+        Should be padded like so:
+
+            [0, 1, 3, 5]
+            [0, 0, 0, 7]
+            [2, 6, 8, 9]
+
+        And ``left_padding`` specifies the amount of padding for each.
+        In this case, ``left_padding = [1, 3, 0]``.
+        """
+
+    def update_and_fetch(self, keys, values):  # -> tuple[array | Any, array | Any]:
+        ...
+    @property
+    def state(
+        self,
+    ):  # -> tuple[Any | array | None, Any | array | None, array | Any, array | Any]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    def is_trimmable(self):  # -> Literal[True]:
+        ...
+    def trim(self, n):  # -> int | float:
+        ...
+    def make_mask(self, N: int, return_array: bool = ..., **kwargs):  # -> array:
+        ...
+    def filter(self, batch_indices):  # -> None:
+        """
+        In-place filter to keep just the given indices in the cache.
+        """
+
+    def extend(self, other):  # -> None:
+        """
+        In-place extend this cache with the other cache.
+        """
+
+class BatchRotatingKVCache(_BaseCache):
+    step = ...
+    def __init__(self, max_size, left_padding: List[int]) -> None: ...
+    def update_and_fetch(
+        self, keys, values
+    ):  # -> tuple[array | Any, array | Any] | tuple[array | Any, array | Any | None]:
+        ...
+    @property
+    def state(
+        self,
+    ):  # -> tuple[Any | array | None, Any | array | None, array | Any, array | Any]:
+        ...
+    @state.setter
+    def state(self, v):  # -> None:
+        ...
+    @property
+    def meta_state(self):  # -> tuple[str, ...]:
+        ...
+    @meta_state.setter
+    def meta_state(self, v):  # -> None:
+        ...
+    def is_trimmable(self):  # -> bool:
+        ...
+    def trim(self, n):  # -> int:
+        ...
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ...
+    ) -> QuantizedKVCache: ...
+    def make_mask(
+        self, N: int, window_size: Optional[int] = ..., return_array: bool = ...
+    ):  # -> array:
+        ...
+    def filter(self, batch_indices):  # -> None:
+        """
+        In-place filter to keep just the given indices in the cache.
+        """
+
+    def extend(self, other):  # -> None:
+        """
+        In-place extend this cache with the other cache.
+        """

.mlx_typings/mlx_lm/models/switch_layers.pyi

@@ -0,0 +1,79 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from functools import partial
+
+import mlx.core as mx
+import mlx.nn as nn
+
+class QuantizedSwitchLinear(nn.Module):
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        num_experts: int,
+        bias: bool = ...,
+        group_size: int = ...,
+        bits: int = ...,
+        mode: str = ...,
+    ) -> None: ...
+    @property
+    def input_dims(self):  # -> int:
+        ...
+    @property
+    def output_dims(self):  # -> int:
+        ...
+    @property
+    def num_experts(self):  # -> int:
+        ...
+    def __call__(self, x, indices, sorted_indices=...):  # -> array:
+        ...
+
+class SwitchLinear(nn.Module):
+    def __init__(
+        self, input_dims: int, output_dims: int, num_experts: int, bias: bool = ...
+    ) -> None: ...
+    @property
+    def input_dims(self):  # -> int:
+        ...
+    @property
+    def output_dims(self):  # -> int:
+        ...
+    @property
+    def num_experts(self):  # -> int:
+        ...
+    def __call__(self, x, indices, sorted_indices=...): ...
+    def to_quantized(
+        self, group_size: int = ..., bits: int = ..., mode: str = ...
+    ):  # -> QuantizedSwitchLinear:
+        ...
+
+@partial(mx.compile, shapeless=True)
+def swiglu(x, gate): ...
+
+class SwiGLU(nn.Module):
+    def __init__(self) -> None: ...
+    def __call__(self, x, gate): ...
+
+class SwitchGLU(nn.Module):
+    def __init__(
+        self,
+        input_dims: int,
+        hidden_dims: int,
+        num_experts: int,
+        activation=...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x, indices) -> mx.array: ...
+
+class SwitchMLP(nn.Module):
+    def __init__(
+        self,
+        input_dims: int,
+        hidden_dims: int,
+        num_experts: int,
+        activation=...,
+        bias: bool = ...,
+    ) -> None: ...
+    def __call__(self, x, indices) -> mx.array: ...

.mlx_typings/mlx_lm/sample_utils.pyi

@@ -0,0 +1,148 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from functools import partial
+from typing import Callable, Dict, List, Optional
+
+import mlx.core as mx
+
+def make_sampler(
+    temp: float = ...,
+    top_p: float = ...,
+    min_p: float = ...,
+    min_tokens_to_keep: int = ...,
+    top_k: int = ...,
+    xtc_probability: float = ...,
+    xtc_threshold: float = ...,
+    xtc_special_tokens: List[int] = ...,
+) -> Callable[[mx.array], mx.array]:
+    """
+    Make a sampler function for use with ``generate_step``.
+
+    Args:
+        temp (float): The temperature for sampling, if 0 the argmax is used.
+          Default: ``0``.
+        top_p (float, optional): Nulceus sampling, higher means model considers
+          more less likely words.
+        min_p (float, optional): The minimum value (scaled by the top token's
+          probability) that a token probability must have to be considered.
+        min_tokens_to_keep (int, optional): Minimum number of tokens that cannot
+          be filtered by min_p sampling.
+        top_k (int, optional): The top k tokens ranked by probability to constrain
+          the sampling to.
+        xtc_probability (float, optional): The probability of applying XTC
+            sampling.
+        xtc_threshold (float, optional): The threshold the probs need to reach
+            for being sampled.
+        xtc_special_tokens (list(int), optional): List of special tokens IDs to
+            be excluded from XTC sampling.
+
+
+    Returns:
+        Callable[mx.array, mx.array]:
+            A sampler which takes log-probabilities and returns tokens.
+    """
+
+def make_logits_processors(
+    logit_bias: Optional[Dict[int, float]] = ...,
+    repetition_penalty: Optional[float] = ...,
+    repetition_context_size: Optional[int] = ...,
+):  # -> list[Any]:
+    """
+    Make logits processors for use with ``generate_step``.
+
+    Args:
+        repetition_penalty (float, optional): The penalty factor for repeating
+          tokens.
+        repetition_context_size (int, optional): The number of tokens to
+          consider for repetition penalty. Default: ``20``.
+        logit_bias (dictionary, optional): Additive logit bias.
+
+    Returns:
+        List[Callable[[mx.array, mx.array], mx.array]]:
+            A list of logits processors. Each processor in the list is a
+            callable which takes an array of tokens and an array of logits
+            and returns the updated logits.
+    """
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_top_k(logprobs: mx.array, top_k: int) -> mx.array:
+    """
+    Sample from only the top K tokens ranked by probability.
+
+    Args:
+        logprobs: A vector of log probabilities.
+        top_k (int): Top k tokens to sample from.
+    """
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_min_p(
+    logprobs: mx.array, min_p: float, min_tokens_to_keep: int = ...
+) -> mx.array:
+    """
+    Apply min-p sampling to the logprobs.
+
+    Min-p keeps all tokens that are above a minimum probability, scaled by the
+    probability of the most likely token. As a result, the filter is more
+    aggressive given a very high-probability token.
+
+    Args:
+        logprobs: A vector of log probabilities.
+        min_p (float): Minimum token probability. Typical values are in the
+            0.01-0.2 range, comparably selective as setting `top_p` in the
+            0.99-0.8 range.
+        min_tokens_to_keep (int, optional): Minimum number of tokens that cannot
+            be filtered. Default: ``1``.
+
+    """
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_top_p(logprobs: mx.array, top_p: float) -> mx.array:
+    """
+    Apply top-p (nucleus) sampling to logits.
+
+    Args:
+        logprobs: A vector of log probabilities.
+        top_p: The cumulative probability threshold for top-p filtering.
+    Returns:
+        token selected based on the top-p criterion.
+    """
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def apply_xtc(
+    logits: mx.array,
+    xtc_probability: float,
+    xtc_threshold: float,
+    xtc_special_tokens: List[int],
+) -> mx.array:
+    """
+    Apply XTC sampling to the logits.
+
+    Args:
+        logits: The logits from the model's output.
+        xtc_probability (float): Probability of XTC sampling to happen for each token
+        xtc_threshold (float): The threshold the probs need to reach for being sampled.
+        special_tokens_ids (list(int)): List of special tokens IDs to be excluded from XTC sampling.
+    """
+
+@partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state)
+def categorical_sampling(logits, temp):  # -> array:
+    ...
+def make_repetition_penalty(
+    penalty: float, context_size: int = ...
+):  # -> Callable[..., Any]:
+    """
+    Make repetition penalty processor.
+
+    Paper: https://arxiv.org/abs/1909.05858
+
+    Args:
+        penalty (float): The repetition penalty factor to be applied.
+        context_size (int): The number of previous tokens to use.
+            Default: ``20``.
+
+    Returns:
+        Callable[[mx.array, List[int]], mx.array]:
+            The repetition penalty processor.
+    """

.mlx_typings/mlx_lm/tokenizer_utils.pyi

@@ -0,0 +1,168 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+from functools import partial
+from pathlib import Path
+
+from transformers import PreTrainedTokenizerFast
+
+class StreamingDetokenizer:
+    """The streaming detokenizer interface so that we can detokenize one token at a time.
+
+    Example usage is as follows:
+
+        detokenizer = ...
+
+        # Reset the tokenizer state
+        detokenizer.reset()
+
+        for token in generate(...):
+            detokenizer.add_token(token.item())
+
+            # Contains the whole text so far. Some tokens may not be included
+            # since it contains whole words usually.
+            detokenizer.text
+
+            # Contains the printable segment (usually a word) since the last
+            # time it was accessed
+            detokenizer.last_segment
+
+            # Contains all the tokens added so far
+            detokenizer.tokens
+
+        # Make sure that we detokenize any remaining tokens
+        detokenizer.finalize()
+
+        # Now detokenizer.text should match tokenizer.decode(detokenizer.tokens)
+    """
+
+    __slots__ = ...
+    def reset(self): ...
+    def add_token(self, token): ...
+    def finalize(self): ...
+    @property
+    def last_segment(self):
+        """Return the last segment of readable text since last time this property was accessed."""
+
+class NaiveStreamingDetokenizer(StreamingDetokenizer):
+    """NaiveStreamingDetokenizer relies on the underlying tokenizer
+    implementation and should work with every tokenizer.
+
+    Its complexity is O(T^2) where T is the longest line since it will
+    repeatedly detokenize the same tokens until a new line is generated.
+    """
+    def __init__(self, tokenizer) -> None: ...
+    def reset(self):  # -> None:
+        ...
+    def add_token(self, token):  # -> None:
+        ...
+    def finalize(self):  # -> None:
+        ...
+    @property
+    def text(self):  # -> str:
+        ...
+
+class SPMStreamingDetokenizer(StreamingDetokenizer):
+    """A streaming detokenizer for SPM models.
+
+    It adds tokens to the text if the next token starts with the special SPM
+    underscore which results in linear complexity.
+    """
+    def __init__(self, tokenizer, trim_space=...) -> None: ...
+    def reset(self):  # -> None:
+        ...
+    def add_token(self, token):  # -> None:
+        ...
+    def finalize(self):  # -> None:
+        ...
+
+class BPEStreamingDetokenizer(StreamingDetokenizer):
+    """A streaming detokenizer for OpenAI style BPE models.
+
+    It adds tokens to the text if the next token starts with a space similar to
+    the SPM detokenizer.
+    """
+
+    _byte_decoder = ...
+    _space_matches = ...
+    def __init__(self, tokenizer) -> None: ...
+    def reset(self):  # -> None:
+        ...
+    def add_token(self, token):  # -> None:
+        ...
+    def finalize(self):  # -> None:
+        ...
+    @classmethod
+    def make_byte_decoder(cls):  # -> None:
+        """See https://github.com/openai/gpt-2/blob/master/src/encoder.py for the rationale."""
+
+class TokenizerWrapper:
+    """A wrapper that combines an HF tokenizer and a detokenizer.
+
+    Accessing any attribute other than the ``detokenizer`` is forwarded to the
+    huggingface tokenizer.
+    """
+    def __init__(self, tokenizer, detokenizer_class=..., eos_token_ids=...) -> None: ...
+    def add_eos_token(self, token: str):  # -> None:
+        ...
+    @property
+    def has_thinking(self):  # -> bool:
+        ...
+    @property
+    def think_start(self):  # -> str | None:
+        ...
+    @property
+    def think_end(self):  # -> str | None:
+        ...
+    @property
+    def has_tool_calling(self):  # -> bool:
+        ...
+    @property
+    def tool_call_start(self):  # -> str | None:
+        ...
+    @property
+    def tool_call_end(self):  # -> str | None:
+        ...
+    @property
+    def detokenizer(self):  # -> NaiveStreamingDetokenizer:
+        """
+        Get a stateful streaming detokenizer.
+        """
+
+    def __getattr__(self, attr):  # -> set[Any] | Any:
+        ...
+    def __setattr__(self, attr, value):  # -> None:
+        ...
+
+class NewlineTokenizer(PreTrainedTokenizerFast):
+    """A tokenizer that replaces newlines with <n> and <n> with new line."""
+    def __init__(self, *args, **kwargs) -> None: ...
+    def encode(self, text, **kwargs):  # -> list[int]:
+        ...
+    def encode_batch(self, texts, **kwargs): ...
+    def decode(self, *args, **kwargs):  # -> str:
+        ...
+    def batch_decode(self, *args, **kwargs):  # -> list[str]:
+        ...
+
+def load_tokenizer(
+    model_path: Path,
+    tokenizer_config_extra=...,
+    return_tokenizer=...,
+    eos_token_ids=...,
+) -> (
+    TokenizerWrapper
+    | type[SPMStreamingDetokenizer]
+    | partial[SPMStreamingDetokenizer]
+    | type[BPEStreamingDetokenizer]
+    | type[NaiveStreamingDetokenizer]
+):
+    """Load a huggingface tokenizer and try to infer the type of streaming
+    detokenizer to use.
+
+    Note, to use a fast streaming tokenizer, pass a local file path rather than
+    a Hugging Face repo ID.
+    """
+
+def no_bos_or_eos(sequence: list, bos: int, eos: int) -> list: ...

.mlx_typings/mlx_lm/utils.pyi

@@ -0,0 +1,195 @@
+"""
+This type stub file was generated by pyright.
+"""
+
+import os
+from pathlib import Path
+from typing import Any, Callable, Dict, Optional, Tuple, Type, Union
+
+import mlx.nn as nn
+from transformers.utils.auto_docstring import ModelArgs
+
+from .tokenizer_utils import TokenizerWrapper
+
+if os.getenv("MLXLM_USE_MODELSCOPE", "False").lower() == "true": ...
+else: ...
+MODEL_REMAPPING = ...
+MAX_FILE_SIZE_GB = ...
+
+def compute_bits_per_weight(model): ...
+def hf_repo_to_path(hf_repo):  # -> Path:
+    ...
+def load_config(model_path: Path) -> dict: ...
+def load_model(
+    model_path: Path,
+    lazy: bool = False,
+    strict: bool = True,
+    model_config: dict[str, Any] = {},
+    get_model_classes: Callable[
+        [dict[str, Any]], Tuple[Type[nn.Module], Type[ModelArgs]]
+    ] = ...,
+) -> Tuple[nn.Module, dict[str, Any]]:
+    """
+    Load and initialize the model from a given path.
+
+    Args:
+        model_path (Path): The path to load the model from.
+        lazy (bool): If False eval the model parameters to make sure they are
+            loaded in memory before returning, otherwise they will be loaded
+            when needed. Default: ``False``
+        strict (bool): Whether or not to raise an exception if weights don't
+            match. Default: ``True``
+        model_config (dict, optional): Optional configuration parameters for the
+            model. Defaults to an empty dictionary.
+        get_model_classes (Callable[[dict], Tuple[Type[nn.Module], Type]], optional):
+            A function that returns the model class and model args class given a config.
+            Defaults to the ``_get_classes`` function.
+
+    Returns:
+        Tuple[nn.Module, dict[str, Any]]: The loaded and initialized model and config.
+
+    Raises:
+        FileNotFoundError: If the weight files (.safetensors) are not found.
+        ValueError: If the model class or args class are not found or cannot be instantiated.
+    """
+
+def load(
+    path_or_hf_repo: str,
+    tokenizer_config=...,
+    model_config=...,
+    adapter_path: Optional[str] = ...,
+    lazy: bool = ...,
+    return_config: bool = ...,
+    revision: str = ...,
+) -> Union[
+    Tuple[nn.Module, TokenizerWrapper],
+    Tuple[nn.Module, TokenizerWrapper, Dict[str, Any]],
+]:
+    """
+    Load the model and tokenizer from a given path or a huggingface repository.
+
+    Args:
+        path_or_hf_repo (Path): The path or the huggingface repository to load the model from.
+        tokenizer_config (dict, optional): Configuration parameters specifically for the tokenizer.
+            Defaults to an empty dictionary.
+        model_config(dict, optional): Configuration parameters specifically for the model.
+            Defaults to an empty dictionary.
+        adapter_path (str, optional): Path to the LoRA adapters. If provided, applies LoRA layers
+            to the model. Default: ``None``.
+        lazy (bool): If ``False`` eval the model parameters to make sure they are
+            loaded in memory before returning, otherwise they will be loaded
+            when needed. Default: ``False``
+        return_config (bool: If ``True`` return the model config as the last item..
+        revision (str, optional): A revision id which can be a branch name, a tag, or a commit hash.
+    Returns:
+        Union[Tuple[nn.Module, TokenizerWrapper], Tuple[nn.Module, TokenizerWrapper, Dict[str, Any]]]:
+            A tuple containing the loaded model, tokenizer and, if requested, the model config.
+
+    Raises:
+        FileNotFoundError: If config file or safetensors are not found.
+        ValueError: If model class or args class are not found.
+    """
+
+def make_shards(weights: dict, max_file_size_gb: int = ...) -> list:
+    """
+    Splits the weights into smaller shards.
+
+    Args:
+        weights (dict): Model weights.
+        max_file_size_gb (int): Maximum size of each shard in gigabytes.
+
+    Returns:
+        list: List of weight shards.
+    """
+
+def create_model_card(
+    path: Union[str, Path], hf_path: Union[str, Path, None]
+):  # -> None:
+    """
+    Uploads the model to Hugging Face hub.
+
+    Args:
+        path (Union[str, Path]): Local path to the model.
+        hf_path (Union[str, Path, None]): Path to the original Hugging Face model.
+    """
+
+def upload_to_hub(path: str, upload_repo: str):  # -> None:
+    """
+    Uploads the model to Hugging Face hub.
+
+    Args:
+        path (str): Local path to the model.
+        upload_repo (str): Name of the HF repo to upload to.
+    """
+
+def save_model(
+    save_path: Union[str, Path], model: nn.Module, *, donate_model: bool = ...
+) -> None:
+    """Save model weights and metadata index into specified directory."""
+
+def quantize_model(
+    model: nn.Module,
+    config: dict,
+    group_size: int,
+    bits: int,
+    mode: str = ...,
+    quant_predicate: Optional[Callable[[str, nn.Module], Union[bool, dict]]] = ...,
+) -> Tuple[nn.Module, dict]:
+    """
+    Applies quantization to the model weights.
+
+    Args:
+        model (nn.Module): The model to be quantized.
+        config (dict): Model configuration.
+        group_size (int): Group size for quantization.
+        bits (int): Bits per weight for quantization.
+        mode (str): The quantization mode.
+        quant_predicate (Callable): A callable that decides how to quantize
+          each layer based on the path. Accepts the layer `path` and the
+          `module`. Returns either a bool to signify quantize/no quantize or
+          a dict of quantization parameters to pass to `to_quantized`.
+
+    Returns:
+        Tuple: Tuple containing quantized model and config.
+    """
+
+def save_config(config: dict, config_path: Union[str, Path]) -> None:
+    """Save the model configuration to the ``config_path``.
+
+    The final configuration will be sorted before saving for better readability.
+
+    Args:
+        config (dict): The model configuration.
+        config_path (Union[str, Path]): Model configuration file path.
+    """
+
+def save(
+    dst_path: Union[str, Path],
+    src_path_or_repo: Union[str, Path],
+    model: nn.Module,
+    tokenizer: TokenizerWrapper,
+    config: Dict[str, Any],
+    donate_model: bool = ...,
+):  # -> None:
+    ...
+def common_prefix_len(list1, list2):  # -> int:
+    """
+    Calculates the length of the common prefix of two lists.
+
+    Args:
+        list1: The first list of strings.
+        list2: The second list of strings.
+
+    Returns:
+        The length of the common prefix. Returns 0 if lists are empty
+        or do not match at the first element.
+    """
+
+def does_model_support_input_embeddings(model: nn.Module) -> bool:
+    """
+    Check if the model supports input_embeddings in its call signature.
+    Args:
+        model (nn.Module): The model to check.
+    Returns:
+        bool: True if the model supports input_embeddings, False otherwise.
+    """

.python-version

@@ -0,0 +1 @@
+3.13

.vscode/extensions.json

@@ -0,0 +1,11 @@
+{
+    "recommendations": [
+        "detachhead.basedpyright",
+        "ms-python.python"
+    ],
+    "unwantedRecommendations": [
+        "ms-python.vscode-pylance",
+        "ms-python.pyright",
+        "ms-python.mypy-type-checker"
+    ]
+}

.vscode/settings.json

@@ -0,0 +1,3 @@
+{
+    "basedpyright.importStrategy": "fromEnvironment"
+}

.zed/settings.json

@@ -0,0 +1,29 @@
+// Folder-specific settings
+//
+// For a full list of overridable settings, and general information on folder-specific settings,
+// see the documentation: https://zed.dev/docs/configuring-zed#settings-files
+{
+  "lsp": {
+    "nix_python": {
+      "binary": {
+        "path": "nix",
+        "arguments": [
+          "run",
+          "--quiet",
+          "--no-warn-dirty",
+          "--no-allow-import-from-derivation",
+          "--print-build-logs",
+          "never",
+          "${projectRoot}#python-lsp",
+          "--",
+          "--stdio"
+        ]
+      }
+    }
+  },
+  "languages": {
+    "Python": {
+      "language_servers": ["nix_python"]
+    }
+  }
+}

Cargo.toml

@@ -0,0 +1,165 @@
+[workspace]
+resolver = "3"
+members = [
+    "rust/networking",
+    "rust/exo_pyo3_bindings",
+    "rust/system_custodian",
+    "rust/util",
+]
+
+[workspace.package]
+version = "0.0.1"
+edition = "2024"
+
+[profile.dev]
+opt-level = 1
+debug = true
+
+[profile.release]
+opt-level = 3
+
+# Common shared dependendencies configured once at the workspace
+# level, to be re-used more easily across workspace member crates.
+#
+# Common configurations include versions, paths, features, etc.
+[workspace.dependencies]
+## Crate members as common dependencies
+networking = { path = "rust/networking" }
+system_custodian = { path = "rust/system_custodian" }
+util = { path = "rust/util" }
+
+# Proc-macro authoring tools
+syn = "2.0"
+quote = "1.0"
+proc-macro2 = "1.0"
+darling = "0.20"
+
+# Macro dependecies
+extend = "1.2"
+delegate = "0.13"
+impl-trait-for-tuples = "0.2"
+clap = "4.5"
+derive_more = { version = "2.0.1", features = ["display"] }
+pin-project = "1"
+
+# Utility dependencies
+itertools = "0.14"
+thiserror = "2"
+internment = "0.8"
+recursion = "0.5"
+regex = "1.11"
+once_cell = "1.21"
+thread_local = "1.1"
+bon = "3.4"
+generativity = "1.1"
+anyhow = "1.0"
+keccak-const = "0.2"
+
+# Functional generics/lenses frameworks
+frunk_core = "0.4"
+frunk = "0.4"
+frunk_utils = "0.2"
+frunk-enum-core = "0.3"
+
+# Async dependencies
+tokio = "1.46"
+futures = "0.3"
+futures-util = "0.3"
+futures-timer = "3.0"
+
+# Data structures
+either = "1.15"
+ordered-float = "5.0"
+ahash = "0.8"
+
+# Tracing/logging
+log = "0.4"
+
+# networking
+libp2p = "0.56"
+libp2p-tcp = "0.44"
+
+[workspace.lints.rust]
+static_mut_refs = "warn"      # Or use "warn" instead of deny
+incomplete_features = "allow"
+
+# Clippy's lint category level configurations;
+# every member crate needs to inherit these by adding
+#
+#     ```toml
+#     [lints]
+#     workspace = true
+#     ```
+#
+# to their `Cargo.toml` files
+[workspace.lints.clippy]
+# Clippy lint categories meant to be enabled all at once
+correctness = { level = "deny", priority = -1 }
+suspicious = { level = "warn", priority = -1 }
+style = { level = "warn", priority = -1 }
+complexity = { level = "warn", priority = -1 }
+perf = { level = "warn", priority = -1 }
+pedantic = { level = "warn", priority = -1 }
+nursery = { level = "warn", priority = -1 }
+cargo = { level = "warn", priority = -1 }
+
+# Individual Clippy lints from the `restriction` category
+arithmetic_side_effects = "warn"
+as_conversions = "warn"
+assertions_on_result_states = "warn"
+clone_on_ref_ptr = "warn"
+decimal_literal_representation = "warn"
+default_union_representation = "warn"
+deref_by_slicing = "warn"
+disallowed_script_idents = "deny"
+else_if_without_else = "warn"
+empty_enum_variants_with_brackets = "warn"
+empty_structs_with_brackets = "warn"
+error_impl_error = "warn"
+exit = "deny"
+expect_used = "warn"
+float_cmp_const = "warn"
+get_unwrap = "warn"
+if_then_some_else_none = "warn"
+impl_trait_in_params = "warn"
+indexing_slicing = "warn"
+infinite_loop = "warn"
+let_underscore_must_use = "warn"
+let_underscore_untyped = "warn"
+lossy_float_literal = "warn"
+mem_forget = "warn"
+missing_inline_in_public_items = "warn"
+multiple_inherent_impl = "warn"
+multiple_unsafe_ops_per_block = "warn"
+mutex_atomic = "warn"
+non_zero_suggestions = "warn"
+panic = "warn"
+partial_pub_fields = "warn"
+pattern_type_mismatch = "warn"
+pub_without_shorthand = "warn"
+rc_buffer = "warn"
+rc_mutex = "warn"
+redundant_type_annotations = "warn"
+renamed_function_params = "warn"
+rest_pat_in_fully_bound_structs = "warn"
+same_name_method = "warn"
+self_named_module_files = "deny"
+semicolon_inside_block = "warn"
+shadow_same = "warn"
+shadow_unrelated = "warn"
+str_to_string = "warn"
+string_add = "warn"
+string_lit_chars_any = "warn"
+string_to_string = "warn"
+tests_outside_test_module = "warn"
+todo = "warn"
+try_err = "warn"
+undocumented_unsafe_blocks = "warn"
+unnecessary_safety_comment = "warn"
+unnecessary_safety_doc = "warn"
+unneeded_field_pattern = "warn"
+unseparated_literal_suffix = "warn"
+unused_result_ok = "warn"
+unused_trait_names = "warn"
+unwrap_used = "warn"
+verbose_file_reads = "warn"

RULES.md

@@ -0,0 +1,84 @@
+# Repository Rules
+
+* if you see any code that violates these rules, raise it with me directly rather than trying to fix.
+  * where applicable, file a GitHub Issue.
+* adhere to these rules strictly.
+
+## General Rules
+
+* if its possible to eliminate an extra try-catch or if-statement at runtime using type-level discipline, do it!
+* name your types, functions, and classes appropriately.
+  * no three-letter acronyms.
+  * no non-standard contractions.
+  * each data type has a meaning, pick a name which is accurate and descriptive.
+  * the average layman should be able to easily understand what your function does using the function signature alone!
+    * sometimes, there will be exceptions. eg, when you're using specific technical terms that are well understood (saga, event, etc).
+    * usually, you'll think that your code is an exception to the rules, but it won't be.
+
+## State, Functions and Classes
+
+* every function, given the same inputs, should produce the same outputs. ie, no hidden state.
+* use classes to prevent fixed state from being mutated arbitrarily (unsafely); methods provide a safe way of interfacing with state.
+* if your logic doesn't mutate fixed state, it probably belongs in a standalone function rather than a class.
+* functions shouldn't usually produce side-effects (they should be computationally pure).
+  * if, for example, you're updating a state using an event (computationally pure), and you want to trigger a saga (computational side-effect), store the logic for triggering the saga into an effect handler (a function, capable of producing side-effects, that you pass into an otherwise computationally pure function, so that it may trigger side-effects safely).
+
+## Pydantic
+
+* read the Pydantic docs.
+* respect the Pydantic docs.
+* pydantic is all you need.
+* declare and re-use a central `ConfigDict` for your use-case, you'll usually want `frozen` and `strict` to be `True`.
+
+## Unique ID (UUID) Generation
+
+* inherit from Pydantic's `UUID4` class to create your own UUID class.
+* use `uuid.uuid4()` to initialize your class with a fresh UUID where possible.
+* ensure that idempotency tags are generated by taking the salted hash of persisted state.
+  * rationale: if a node crashes and resumes from an older state, it should not accidentally re-publish the same event twice under different idempotency tags. 
+  * every distinct function should feature a unique salt, so that there are no accidental collisions in idempotency tags.
+
+## Type Wrappers
+
+* reuse types that already exist in the Python standard library.
+* when two distinct data types are structurally identical (for example, different IDs which are both UUIDs but shouldn't never mixed up), make sure they can't be conflated by the type system.
+  * if you're working with a primitive data type (`str`, `int`, etc), use `NewType` (it has zero runtime overhead).
+  * if you're working with serializable data objects, consider adding a field (type `str`) that states its type.
+
+## Type Discipline
+
+* do not bypass the type-checker, preserve strict typing by any means necessary.
+* by default, use literal types (like `Literal['one', 'two']`) where an enum seems appropriate.
+
+pro-tip: Python's type system is quite complex and feature-rich, so reading the documentation is often advisable; Matt discovered that Python `typing` library allows you to check that you've implemented a `match` exhaustively using `Literal` and `get_args(type)` after reading the docs.
+
+## Use of `@final`, Freezing
+
+* use wherever applicable.
+
+## Error Handling
+
+* don't try-catch for no reason.
+* make sure that you always know where and when the exceptions your code produces are meant to be handled, so that it's never a nasty surprise.
+  * always write the rationale for your error-handling down in the docstring!
+  * communicate the details to your colleagues when appropriate.
+
+## Dependencies
+
+* don't introduce any new dependencies without asking.
+* don't ask for any dependencies that aren't ubiquitous within production environments.
+
+## Commit Messages
+
+*   use the imperative mood in the subject line.
+*   prefix the subject line with a change type. our change types are:
+    *   `documentation`: documentation changes.
+    *   `feature`: a new feature.
+    *   `refactor`: a code change that neither fixes a bug nor adds a feature.
+    *   `bugfix`: a bug fix.
+    *   `chore`: routine tasks, maintenance, or tooling changes.
+    *   `test`: adding or correcting tests.
+*   restrict the subject line to fifty characters or less.
+*   capitalize the subject line.
+*   do not end the subject line with a period.
+*   separate subject from body with a blank line.

TODO.md

@@ -0,0 +1,27 @@
+2. Currently a lot of requests from the API are timing out, but we still process those requests internally. If an API request times out, we should cancel all corresponding tasks to that API request (why process a request with nobody listening).
+3. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
+4. I'd like to see profiled network latency / bandwidth.
+5. I'd like to see how much bandwidth each link is using.
+6. We should handle the case where one machine doesn't have the model downloaded and then other machines are waiting on it. In this case we get loads of timeout errors because the others are waiting for the one that needs to download the model.
+7. Solve the problem of in continuous batching when a new prompt comes in, it will block decode of the current batch until the prefill is complete.
+8. We want people to be able to copy models over to a new device without ever connecting EXO to the internet. Right now EXO require internet connection once to cache some files to check if a download is complete. Instead, we should simply check if there is a non-empty model folder locally with no .partial files. This indicates it's a fully downloaded model that can be loaded.
+10. More granular control over how to deploy instances.
+12. Nix is great but installing it is a pain and we have ended up in a lot of cases having PATH issues or installation issues. For example, after rebooting mike it seemed to no longer have a nix installation and needed reinstalling. It has a bunch of broken symlinks left over from nix that caused ssh to fail, making it even harder to debug. We need consistent environments (perhaps MDM) so we can guarantee nix is installed properly on each machine.
+13. Memory pressure instead of memory used.
+14. Show the type of each connection (TB5, Ethernet, etc.) in the UI. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
+15. Prioritise certain connection types (or by latency). TB5 > Ethernet > WiFi. Refer to old exo: https://github.com/exo-explore/exo/blob/56f783b38dc6b08ce606b07a5386dc40dae00330/exo/helpers.py#L251
+16. Dynamically switch to higher priority connection when it becomes available. Probably bring back InstanceReplacedAtomically.
+17. Faster model loads by streaming model from other devices in cluster.
+18. Add support for specifying the type of network connection to use in a test. Depends on 15/16.
+20. Add chat completion cancellations (e.g OpenWebUI has something for cancelling an ongoing request).
+23. Do we need cache_limit? We went back and forth on that a lot because we thought it might be causing issues. One problem is it sets it relative to model size. So if you have multiple models loaded in it will take the most recent model size for the cache_limit. This is problematic if you launch DeepSeek -> Llama for example.
+24. further openai/lmstudio api compatibility
+25. Rethink retry logic
+26. Task cancellation. When API http request gets cancelled, it should cancel corresponding task.
+27. Log cleanup - per-module log filters and default to DEBUG log levels
+
+Potential refactors:
+
+2. Topology can be simplified
+
+Random errors we've run into:

dashboard/package.json

@@ -0,0 +1,33 @@
+{
+	"name": "exo-dashboard",
+	"private": true,
+	"version": "1.0.0",
+	"type": "module",
+	"scripts": {
+		"dev": "vite dev",
+		"build": "vite build",
+		"preview": "vite preview",
+		"prepare": "svelte-kit sync || echo ''",
+		"check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json"
+	},
+	"devDependencies": {
+		"@sveltejs/adapter-static": "^3.0.10",
+		"@sveltejs/kit": "^2.48.4",
+		"@sveltejs/vite-plugin-svelte": "^5.0.0",
+		"@tailwindcss/vite": "^4.0.0",
+		"@types/d3": "^7.4.3",
+		"@types/node": "^22",
+		"d3": "^7.9.0",
+		"svelte": "^5.0.0",
+		"svelte-check": "^4.0.0",
+		"tailwindcss": "^4.0.0",
+		"tw-animate-css": "^1.3.5",
+		"typescript": "^5.0.0",
+		"vite": "^6.0.0"
+	},
+	"dependencies": {
+		"highlight.js": "^11.11.1",
+		"mode-watcher": "^1.1.0"
+	}
+}
+

dashboard/src/app.css

@@ -0,0 +1,322 @@
+@import 'tailwindcss';
+@import 'tw-animate-css';
+
+@custom-variant dark (&:is(.dark *));
+
+:root {
+	/* EXO Brand Colors - Command Center Theme (neutral dark greys) */
+	--exo-black: oklch(0.12 0 0);
+	--exo-dark-gray: oklch(0.16 0 0);
+	--exo-medium-gray: oklch(0.22 0 0);
+	--exo-light-gray: oklch(0.6 0 0);
+	--exo-yellow: oklch(0.85 0.18 85);
+	--exo-yellow-darker: oklch(0.7 0.16 85);
+	--exo-yellow-glow: oklch(0.9 0.2 85);
+	
+	/* Gotham-inspired accent colors */
+	--exo-grid: oklch(0.25 0 0);
+	--exo-scanline: oklch(0.15 0 0);
+	--exo-glow-yellow: 0 0 20px oklch(0.85 0.18 85 / 0.3);
+	--exo-glow-yellow-strong: 0 0 40px oklch(0.85 0.18 85 / 0.5);
+	
+	/* Theme Variables */
+	--radius: 0.375rem;
+	--background: var(--exo-black);
+	--foreground: oklch(0.9 0 0);
+	--card: var(--exo-dark-gray);
+	--card-foreground: oklch(0.9 0 0);
+	--popover: var(--exo-dark-gray);
+	--popover-foreground: oklch(0.9 0 0);
+	--primary: var(--exo-yellow);
+	--primary-foreground: var(--exo-black);
+	--secondary: var(--exo-medium-gray);
+	--secondary-foreground: oklch(0.9 0 0);
+	--muted: var(--exo-medium-gray);
+	--muted-foreground: var(--exo-light-gray);
+	--accent: var(--exo-medium-gray);
+	--accent-foreground: oklch(0.9 0 0);
+	--destructive: oklch(0.6 0.25 25);
+	--border: oklch(0.22 0 0);
+	--input: oklch(0.22 0 0);
+	--ring: var(--exo-yellow);
+}
+
+@theme inline {
+	--radius-sm: calc(var(--radius) - 2px);
+	--radius-md: var(--radius);
+	--radius-lg: calc(var(--radius) + 2px);
+	--radius-xl: calc(var(--radius) + 4px);
+	--color-background: var(--background);
+	--color-foreground: var(--foreground);
+	--color-card: var(--card);
+	--color-card-foreground: var(--card-foreground);
+	--color-popover: var(--popover);
+	--color-popover-foreground: var(--popover-foreground);
+	--color-primary: var(--primary);
+	--color-primary-foreground: var(--primary-foreground);
+	--color-secondary: var(--secondary);
+	--color-secondary-foreground: var(--secondary-foreground);
+	--color-muted: var(--muted);
+	--color-muted-foreground: var(--muted-foreground);
+	--color-accent: var(--accent);
+	--color-accent-foreground: var(--accent-foreground);
+	--color-destructive: var(--destructive);
+	--color-border: var(--border);
+	--color-input: var(--input);
+	--color-ring: var(--ring);
+	
+	/* Custom EXO colors */
+	--color-exo-yellow: var(--exo-yellow);
+	--color-exo-yellow-darker: var(--exo-yellow-darker);
+	--color-exo-black: var(--exo-black);
+	--color-exo-dark-gray: var(--exo-dark-gray);
+	--color-exo-medium-gray: var(--exo-medium-gray);
+	--color-exo-light-gray: var(--exo-light-gray);
+}
+
+@layer base {
+	* {
+		@apply border-border outline-ring/50;
+	}
+	html, body {
+		@apply bg-background text-foreground;
+		font-family: 'SF Mono', 'Fira Code', 'Monaco', 'Consolas', 'Liberation Mono', monospace;
+		letter-spacing: 0.02em;
+	}
+}
+
+@layer utilities {
+	.scrollbar-hide {
+		&::-webkit-scrollbar {
+			display: none;
+		}
+		-ms-overflow-style: none;
+		scrollbar-width: none;
+	}
+	
+	/* CRT Scanline effect */
+	.scanlines {
+		position: relative;
+		&::before {
+			content: '';
+			position: absolute;
+			inset: 0;
+			background: repeating-linear-gradient(
+				0deg,
+				transparent,
+				transparent 2px,
+				oklch(0 0 0 / 0.03) 2px,
+				oklch(0 0 0 / 0.03) 4px
+			);
+			pointer-events: none;
+			z-index: 100;
+		}
+	}
+	
+	/* Command panel styling */
+	.command-panel {
+		background: linear-gradient(
+			180deg,
+			oklch(0.16 0 0 / 0.95) 0%,
+			oklch(0.12 0 0 / 0.98) 100%
+		);
+		border: 1px solid oklch(0.25 0 0);
+		box-shadow: 
+			inset 0 1px 0 oklch(1 0 0 / 0.03),
+			0 4px 20px oklch(0 0 0 / 0.5);
+	}
+	
+	/* Glow text */
+	.glow-text {
+		text-shadow: 
+			0 0 10px oklch(0.85 0.18 85 / 0.5),
+			0 0 20px oklch(0.85 0.18 85 / 0.3),
+			0 0 40px oklch(0.85 0.18 85 / 0.1);
+	}
+	
+	/* Status indicator pulse */
+	.status-pulse {
+		animation: statusPulse 2s ease-in-out infinite;
+	}
+	
+	/* Grid background */
+	.grid-bg {
+		background-image: 
+			linear-gradient(oklch(0.2 0 0 / 0.3) 1px, transparent 1px),
+			linear-gradient(90deg, oklch(0.2 0 0 / 0.3) 1px, transparent 1px);
+		background-size: 40px 40px;
+	}
+}
+
+/* Animations */
+@keyframes flowAnimation {
+	from {
+		stroke-dashoffset: 0;
+	}
+	to {
+		stroke-dashoffset: -16;
+	}
+}
+
+@keyframes statusPulse {
+	0%, 100% {
+		opacity: 1;
+	}
+	50% {
+		opacity: 0.5;
+	}
+}
+
+@keyframes radarSweep {
+	from {
+		transform: rotate(0deg);
+	}
+	to {
+		transform: rotate(360deg);
+	}
+}
+
+@keyframes glowPulse {
+	0%, 100% {
+		box-shadow: 0 0 5px oklch(0.85 0.18 85 / 0.3), 0 0 10px oklch(0.85 0.18 85 / 0.1);
+	}
+	50% {
+		box-shadow: 0 0 15px oklch(0.85 0.18 85 / 0.5), 0 0 30px oklch(0.85 0.18 85 / 0.2);
+	}
+}
+
+@keyframes dataPulse {
+	0%, 100% {
+		opacity: 0.6;
+	}
+	50% {
+		opacity: 1;
+	}
+}
+
+.graph-link {
+	stroke: oklch(0.85 0.18 85 / 0.4);
+	stroke-width: 1.5px;
+	stroke-dasharray: 8, 8;
+	animation: flowAnimation 1s linear infinite;
+	filter: drop-shadow(0 0 3px oklch(0.85 0.18 85 / 0.5));
+}
+
+.graph-link-active {
+	stroke: oklch(0.85 0.18 85 / 0.8);
+	stroke-width: 2px;
+	filter: drop-shadow(0 0 6px oklch(0.85 0.18 85 / 0.8));
+}
+
+/* CRT Screen effect for topology */
+.crt-screen {
+	position: relative;
+	border-radius: 50%;
+	background: radial-gradient(
+		ellipse at center,
+		oklch(0.16 0 0) 0%,
+		oklch(0.12 0 0) 50%,
+		oklch(0.09 0 0) 100%
+	);
+	box-shadow:
+		inset 0 0 100px oklch(0 0 0 / 0.5),
+		0 0 50px oklch(0.85 0.18 85 / 0.1);
+}
+
+/* Data readout styling */
+.data-readout {
+	font-family: 'SF Mono', 'Fira Code', monospace;
+	font-size: 11px;
+	letter-spacing: 0.05em;
+	text-transform: uppercase;
+}
+
+/* Terminal cursor blink */
+.cursor-blink {
+	animation: cursorBlink 1s step-end infinite;
+}
+
+@keyframes cursorBlink {
+	0%, 100% { opacity: 1; }
+	50% { opacity: 0; }
+}
+
+/* Custom scrollbar for command center */
+::-webkit-scrollbar {
+	width: 6px;
+	height: 6px;
+}
+
+::-webkit-scrollbar-track {
+	background: oklch(0.1 0 0);
+}
+
+::-webkit-scrollbar-thumb {
+	background: oklch(0.3 0 0);
+	border-radius: 3px;
+}
+
+::-webkit-scrollbar-thumb:hover {
+	background: oklch(0.85 0.18 85 / 0.5);
+}
+
+/* Remove focus outline/border for inputs */
+input:focus, textarea:focus {
+	outline: none;
+	box-shadow: none;
+}
+
+/* Shooting Stars Animation */
+.shooting-stars {
+	position: fixed;
+	inset: 0;
+	overflow: hidden;
+	pointer-events: none;
+	z-index: 0;
+}
+
+.shooting-star {
+	position: absolute;
+	width: 3px;
+	height: 3px;
+	background: oklch(0.85 0.18 85 / 1);
+	border-radius: 50%;
+	box-shadow: 0 0 6px oklch(0.85 0.18 85 / 0.8);
+	animation: shootingStar var(--duration, 3s) linear infinite;
+	animation-delay: var(--delay, 0s);
+	opacity: 0;
+}
+
+.shooting-star::before {
+	content: '';
+	position: absolute;
+	width: 80px;
+	height: 2px;
+	background: linear-gradient(90deg, oklch(0.85 0.18 85 / 0), oklch(0.85 0.18 85 / 0.6));
+	transform: rotate(45deg);
+	transform-origin: right center;
+	top: 0;
+	right: 2px;
+}
+
+@keyframes shootingStar {
+	0% {
+		opacity: 0;
+		transform: translate(0, 0);
+	}
+	0.5% {
+		opacity: 1;
+	}
+	2.5% {
+		opacity: 0.8;
+		transform: translate(300px, 300px);
+	}
+	3.5% {
+		opacity: 0;
+		transform: translate(400px, 400px);
+	}
+	100% {
+		opacity: 0;
+		transform: translate(400px, 400px);
+	}
+}

dashboard/src/app.d.ts

@@ -0,0 +1,14 @@
+// See https://svelte.dev/docs/kit/types#app.d.ts
+// for information about these interfaces
+declare global {
+	namespace App {
+		// interface Error {}
+		// interface Locals {}
+		// interface PageData {}
+		// interface PageState {}
+		// interface Platform {}
+	}
+}
+
+export {};
+

dashboard/src/app.html

@@ -0,0 +1,14 @@
+<!doctype html>
+<html lang="en">
+	<head>
+		<meta charset="utf-8" />
+		<link rel="icon" href="%sveltekit.assets%/favicon.ico" />
+		<meta name="viewport" content="width=device-width, initial-scale=1" />
+		<title>EXO</title>
+		%sveltekit.head%
+	</head>
+	<body data-sveltekit-preload-data="hover">
+		<div style="display: contents">%sveltekit.body%</div>
+	</body>
+</html>
+

dashboard/src/lib/components/ChatAttachments.svelte

@@ -0,0 +1,75 @@
+<script lang="ts">
+	import type { ChatUploadedFile } from '$lib/types/files';
+	import { formatFileSize, getFileCategory } from '$lib/types/files';
+
+	interface Props {
+		files: ChatUploadedFile[];
+		readonly?: boolean;
+		onRemove?: (fileId: string) => void;
+	}
+
+	let { files, readonly = false, onRemove }: Props = $props();
+
+	function getFileIcon(file: ChatUploadedFile): string {
+		const category = getFileCategory(file.type, file.name);
+		switch (category) {
+			case 'image': return '🖼';
+			case 'text': return '📄';
+			case 'pdf': return '📑';
+			case 'audio': return '🎵';
+			default: return '📎';
+		}
+	}
+
+	function truncateName(name: string, maxLen: number = 20): string {
+		if (name.length <= maxLen) return name;
+		const ext = name.slice(name.lastIndexOf('.'));
+		const base = name.slice(0, name.lastIndexOf('.'));
+		const available = maxLen - ext.length - 3;
+		return base.slice(0, available) + '...' + ext;
+	}
+</script>
+
+{#if files.length > 0}
+	<div class="flex flex-wrap gap-2 mb-3 px-1">
+		{#each files as file (file.id)}
+			<div class="group relative flex items-center gap-2 bg-exo-dark-gray/80 border border-exo-yellow/30 rounded px-2.5 py-1.5 text-xs font-mono transition-all hover:border-exo-yellow/50 hover:shadow-[0_0_10px_rgba(255,215,0,0.1)]">
+				<!-- File preview or icon -->
+				{#if file.preview && getFileCategory(file.type, file.name) === 'image'}
+					<img 
+						src={file.preview} 
+						alt={file.name}
+						class="w-8 h-8 object-cover rounded border border-exo-yellow/20"
+					/>
+				{:else}
+					<span class="text-base">{getFileIcon(file)}</span>
+				{/if}
+				
+				<!-- File info -->
+				<div class="flex flex-col min-w-0">
+					<span class="text-exo-yellow truncate max-w-[120px]" title={file.name}>
+						{truncateName(file.name)}
+					</span>
+					<span class="text-exo-light-gray text-xs">
+						{formatFileSize(file.size)}
+					</span>
+				</div>
+				
+				<!-- Remove button -->
+				{#if !readonly && onRemove}
+					<button
+						type="button"
+						onclick={() => onRemove?.(file.id)}
+						class="ml-1 w-4 h-4 flex items-center justify-center text-exo-light-gray hover:text-red-400 transition-colors cursor-pointer"
+						title="Remove file"
+					>
+						<svg class="w-3 h-3" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+							<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M6 18L18 6M6 6l12 12" />
+						</svg>
+					</button>
+				{/if}
+			</div>
+		{/each}
+	</div>
+{/if}
+

dashboard/src/lib/components/ChatForm.svelte

@@ -0,0 +1,398 @@
+<script lang="ts">
+	import { isLoading, sendMessage, selectedChatModel, setSelectedChatModel, instances, ttftMs, tps, totalTokens } from '$lib/stores/app.svelte';
+	import ChatAttachments from './ChatAttachments.svelte';
+	import type { ChatUploadedFile } from '$lib/types/files';
+	import { processUploadedFiles, getAcceptString } from '$lib/types/files';
+
+	interface Props {
+		class?: string;
+		placeholder?: string;
+		showHelperText?: boolean;
+		autofocus?: boolean;
+		showModelSelector?: boolean;
+	}
+
+	let { 
+		class: className = '', 
+		placeholder = 'Ask anything',
+		showHelperText = false,
+		autofocus = true,
+		showModelSelector = false
+	}: Props = $props();
+
+	let message = $state('');
+	let textareaRef: HTMLTextAreaElement | undefined = $state();
+	let fileInputRef: HTMLInputElement | undefined = $state();
+	let uploadedFiles = $state<ChatUploadedFile[]>([]);
+	let isDragOver = $state(false);
+	let loading = $derived(isLoading());
+	const currentModel = $derived(selectedChatModel());
+	const instanceData = $derived(instances());
+	const currentTtft = $derived(ttftMs());
+	const currentTps = $derived(tps());
+	const currentTokens = $derived(totalTokens());
+	
+	// Custom dropdown state
+	let isModelDropdownOpen = $state(false);
+	let dropdownButtonRef: HTMLButtonElement | undefined = $state();
+	let dropdownPosition = $derived(() => {
+		if (!dropdownButtonRef || !isModelDropdownOpen) return { top: 0, left: 0, width: 0 };
+		const rect = dropdownButtonRef.getBoundingClientRect();
+		return {
+			top: rect.top,
+			left: rect.left,
+			width: rect.width
+		};
+	});
+
+	// Accept all supported file types
+	const acceptString = getAcceptString(['image', 'text', 'pdf']);
+
+	// Extract available models from running instances
+	const availableModels = $derived(() => {
+		const models: Array<{id: string, label: string}> = [];
+		for (const [, instance] of Object.entries(instanceData)) {
+			const modelId = getInstanceModelId(instance);
+			if (modelId && modelId !== 'Unknown' && !models.some(m => m.id === modelId)) {
+				models.push({ id: modelId, label: modelId.split('/').pop() || modelId });
+			}
+		}
+		return models;
+	});
+
+	// Auto-select the first available model if none is selected
+	$effect(() => {
+		const models = availableModels();
+		if (models.length > 0 && !currentModel) {
+			setSelectedChatModel(models[0].id);
+		}
+	});
+
+	function getInstanceModelId(instanceWrapped: unknown): string {
+		if (!instanceWrapped || typeof instanceWrapped !== 'object') return '';
+		const keys = Object.keys(instanceWrapped as Record<string, unknown>);
+		if (keys.length === 1) {
+			const instance = (instanceWrapped as Record<string, unknown>)[keys[0]] as { shardAssignments?: { modelId?: string } };
+			return instance?.shardAssignments?.modelId || '';
+		}
+		return '';
+	}
+
+	async function handleFiles(files: File[]) {
+		if (files.length === 0) return;
+		const processed = await processUploadedFiles(files);
+		uploadedFiles = [...uploadedFiles, ...processed];
+	}
+
+	function handleFileInputChange(event: Event) {
+		const input = event.target as HTMLInputElement;
+		if (input.files && input.files.length > 0) {
+			handleFiles(Array.from(input.files));
+			input.value = ''; // Reset for next selection
+		}
+	}
+
+	function handleFileRemove(fileId: string) {
+		uploadedFiles = uploadedFiles.filter(f => f.id !== fileId);
+	}
+
+	function handlePaste(event: ClipboardEvent) {
+		if (!event.clipboardData) return;
+		
+		const files = Array.from(event.clipboardData.items)
+			.filter(item => item.kind === 'file')
+			.map(item => item.getAsFile())
+			.filter((file): file is File => file !== null);
+		
+		if (files.length > 0) {
+			event.preventDefault();
+			handleFiles(files);
+			return;
+		}
+		
+		// Handle long text paste as file
+		const text = event.clipboardData.getData('text/plain');
+		if (text.length > 2500) {
+			event.preventDefault();
+			const textFile = new File([text], 'pasted-text.txt', { type: 'text/plain' });
+			handleFiles([textFile]);
+		}
+	}
+
+	function handleDragOver(event: DragEvent) {
+		event.preventDefault();
+		isDragOver = true;
+	}
+
+	function handleDragLeave(event: DragEvent) {
+		event.preventDefault();
+		isDragOver = false;
+	}
+
+	function handleDrop(event: DragEvent) {
+		event.preventDefault();
+		isDragOver = false;
+		
+		if (event.dataTransfer?.files) {
+			handleFiles(Array.from(event.dataTransfer.files));
+		}
+	}
+
+	function handleKeydown(event: KeyboardEvent) {
+		if (event.key === 'Enter' && !event.shiftKey) {
+			event.preventDefault();
+			handleSubmit();
+		}
+	}
+
+	function handleSubmit() {
+		if ((!message.trim() && uploadedFiles.length === 0) || loading) return;
+		
+		const content = message.trim();
+		const files = [...uploadedFiles];
+		
+		message = '';
+		uploadedFiles = [];
+		resetTextareaHeight();
+		
+		sendMessage(content, files);
+		
+		// Refocus the textarea after sending
+		setTimeout(() => textareaRef?.focus(), 10);
+	}
+
+	function handleInput() {
+		if (!textareaRef) return;
+		textareaRef.style.height = 'auto';
+		textareaRef.style.height = Math.min(textareaRef.scrollHeight, 150) + 'px';
+	}
+
+	function resetTextareaHeight() {
+		if (textareaRef) {
+			textareaRef.style.height = 'auto';
+		}
+	}
+
+	function openFilePicker() {
+		fileInputRef?.click();
+	}
+
+	// Track previous loading state to detect when loading completes
+	let wasLoading = $state(false);
+	
+	$effect(() => {
+		if (autofocus && textareaRef) {
+			setTimeout(() => textareaRef?.focus(), 10);
+		}
+	});
+	
+	// Refocus after loading completes (AI response finished)
+	$effect(() => {
+		if (wasLoading && !loading && textareaRef) {
+			setTimeout(() => textareaRef?.focus(), 50);
+		}
+		wasLoading = loading;
+	});
+
+	const canSend = $derived(message.trim().length > 0 || uploadedFiles.length > 0);
+</script>
+
+<!-- Hidden file input -->
+<input
+	bind:this={fileInputRef}
+	type="file"
+	accept={acceptString}
+	multiple
+	class="hidden"
+	onchange={handleFileInputChange}
+/>
+
+<form 
+	onsubmit={(e) => { e.preventDefault(); handleSubmit(); }} 
+	class="w-full {className}"
+	ondragover={handleDragOver}
+	ondragleave={handleDragLeave}
+	ondrop={handleDrop}
+>
+	<div 
+		class="relative command-panel rounded overflow-hidden transition-all duration-200 {isDragOver ? 'ring-2 ring-exo-yellow ring-opacity-50' : ''}"
+	>
+		<!-- Top accent line -->
+		<div class="absolute top-0 left-0 right-0 h-px bg-gradient-to-r from-transparent via-exo-yellow/50 to-transparent"></div>
+		
+		<!-- Drag overlay -->
+		{#if isDragOver}
+			<div class="absolute inset-0 bg-exo-dark-gray/80 z-10 flex items-center justify-center">
+				<div class="text-exo-yellow text-sm font-mono tracking-wider uppercase">
+					DROP FILES HERE
+				</div>
+			</div>
+		{/if}
+		
+		<!-- Model selector (when enabled) -->
+		{#if showModelSelector && availableModels().length > 0}
+			<div class="flex items-center justify-between gap-2 px-3 py-2 border-b border-exo-medium-gray/30">
+				<div class="flex items-center gap-2 flex-1">
+					<span class="text-xs text-exo-light-gray uppercase tracking-wider flex-shrink-0">MODEL:</span>
+					<!-- Custom dropdown -->
+					<div class="relative flex-1 max-w-xs">
+						<button
+							bind:this={dropdownButtonRef}
+							type="button"
+							onclick={() => isModelDropdownOpen = !isModelDropdownOpen}
+							class="w-full bg-exo-medium-gray/50 border border-exo-yellow/30 rounded pl-3 pr-8 py-1.5 text-xs font-mono text-left tracking-wide cursor-pointer transition-all duration-200 hover:border-exo-yellow/50 focus:outline-none focus:border-exo-yellow/70 {isModelDropdownOpen ? 'border-exo-yellow/70' : ''}"
+						>
+							{#if availableModels().find(m => m.id === currentModel)}
+								<span class="text-exo-yellow truncate">{availableModels().find(m => m.id === currentModel)?.label}</span>
+							{:else if availableModels().length > 0}
+								<span class="text-exo-yellow truncate">{availableModels()[0].label}</span>
+							{:else}
+								<span class="text-exo-light-gray/50">— SELECT MODEL —</span>
+							{/if}
+						</button>
+						<div class="absolute right-2 top-1/2 -translate-y-1/2 pointer-events-none transition-transform duration-200 {isModelDropdownOpen ? 'rotate-180' : ''}">
+							<svg class="w-3 h-3 text-exo-yellow/60" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+								<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 9l-7 7-7-7" />
+							</svg>
+						</div>
+					</div>
+					
+					{#if isModelDropdownOpen}
+						<!-- Backdrop to close dropdown -->
+						<button 
+							type="button"
+							class="fixed inset-0 z-[9998] cursor-default" 
+							onclick={() => isModelDropdownOpen = false}
+							aria-label="Close dropdown"
+						></button>
+						
+						<!-- Dropdown Panel - fixed positioning to escape overflow:hidden -->
+						<div 
+							class="fixed bg-exo-dark-gray border border-exo-yellow/30 rounded shadow-lg shadow-black/50 z-[9999] max-h-48 overflow-y-auto"
+							style="bottom: calc(100vh - {dropdownPosition().top}px + 4px); left: {dropdownPosition().left}px; width: {dropdownPosition().width}px;"
+						>
+							<div class="py-1">
+								{#each availableModels() as model}
+									<button
+										type="button"
+										onclick={() => {
+											setSelectedChatModel(model.id);
+											isModelDropdownOpen = false;
+										}}
+										class="w-full px-3 py-2 text-left text-xs font-mono tracking-wide transition-colors duration-100 flex items-center gap-2 {
+											currentModel === model.id 
+											? 'bg-transparent text-exo-yellow' 
+											: 'text-exo-light-gray hover:text-exo-yellow'
+										}"
+									>
+										{#if currentModel === model.id}
+											<svg class="w-3 h-3 flex-shrink-0" fill="currentColor" viewBox="0 0 20 20">
+												<path fill-rule="evenodd" d="M16.707 5.293a1 1 0 010 1.414l-8 8a1 1 0 01-1.414 0l-4-4a1 1 0 011.414-1.414L8 12.586l7.293-7.293a1 1 0 011.414 0z" clip-rule="evenodd" />
+											</svg>
+										{:else}
+											<span class="w-3"></span>
+										{/if}
+										<span class="truncate">{model.label}</span>
+									</button>
+								{/each}
+							</div>
+						</div>
+					{/if}
+				</div>
+				<!-- Performance stats -->
+				{#if currentTtft !== null || currentTps !== null}
+					<div class="flex items-center gap-4 text-xs font-mono flex-shrink-0">
+						{#if currentTtft !== null}
+							<span class="text-exo-light-gray">
+								<span class="text-white/70">TTFT</span> <span class="text-exo-yellow">{currentTtft.toFixed(1)}ms</span>
+							</span>
+						{/if}
+						{#if currentTps !== null}
+							<span class="text-exo-light-gray">
+								<span class="text-white/70">TPS</span> <span class="text-exo-yellow">{currentTps.toFixed(1)}</span> <span class="text-white/60">tok/s</span>
+								<span class="text-white/50">({(1000 / currentTps).toFixed(1)} ms/tok)</span>
+							</span>
+						{/if}
+					</div>
+				{/if}
+			</div>
+		{/if}
+		
+		<!-- Attached files preview -->
+		{#if uploadedFiles.length > 0}
+			<div class="px-3 pt-3">
+				<ChatAttachments 
+					files={uploadedFiles} 
+					onRemove={handleFileRemove}
+				/>
+			</div>
+		{/if}
+		
+		<!-- Input area -->
+		<div class="flex items-start gap-2 sm:gap-3 py-3 px-3 sm:px-4">
+			<!-- Attach file button -->
+			<button
+				type="button"
+				onclick={openFilePicker}
+				disabled={loading}
+				class="flex items-center justify-center w-7 h-7 rounded text-exo-light-gray hover:text-exo-yellow transition-all disabled:opacity-50 disabled:cursor-not-allowed flex-shrink-0 cursor-pointer"
+				title="Attach file"
+			>
+				<svg class="w-4 h-4" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+					<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15.172 7l-6.586 6.586a2 2 0 102.828 2.828l6.414-6.586a4 4 0 00-5.656-5.656l-6.415 6.585a6 6 0 108.486 8.486L20.5 13" />
+				</svg>
+			</button>
+			
+			<!-- Terminal prompt -->
+			<span class="text-exo-yellow text-sm font-bold flex-shrink-0 leading-7">▶</span>
+			
+			<textarea
+				bind:this={textareaRef}
+				bind:value={message}
+				onkeydown={handleKeydown}
+				oninput={handleInput}
+				onpaste={handlePaste}
+				{placeholder}
+				disabled={loading}
+				rows={1}
+				class="flex-1 resize-none bg-transparent text-foreground placeholder:text-exo-light-gray/60 placeholder:text-sm placeholder:tracking-[0.15em] placeholder:leading-7 focus:outline-none focus:ring-0 focus:border-none disabled:opacity-50 text-sm leading-7 font-mono"
+				style="min-height: 28px; max-height: 150px;"
+			></textarea>
+			
+			<button
+				type="submit"
+				disabled={!canSend || loading}
+				class="px-2.5 sm:px-4 py-1.5 sm:py-2 rounded text-xs sm:text-xs tracking-[0.1em] sm:tracking-[0.15em] uppercase font-medium transition-all duration-200 whitespace-nowrap
+					{!canSend || loading 
+						? 'bg-exo-medium-gray/50 text-exo-light-gray cursor-not-allowed' 
+						: 'bg-exo-yellow text-exo-black hover:bg-exo-yellow-darker hover:shadow-[0_0_20px_rgba(255,215,0,0.3)]'}"
+				aria-label="Send message"
+			>
+				{#if loading}
+					<span class="inline-flex items-center gap-1 sm:gap-2">
+						<span class="w-2.5 h-2.5 sm:w-3 sm:h-3 border-2 border-current border-t-transparent rounded-full animate-spin"></span>
+						<span class="hidden sm:inline">PROCESSING</span>
+						<span class="sm:hidden">...</span>
+					</span>
+				{:else}
+					SEND
+				{/if}
+			</button>
+		</div>
+		
+		<!-- Bottom accent line -->
+		<div class="absolute bottom-0 left-0 right-0 h-px bg-gradient-to-r from-transparent via-exo-yellow/30 to-transparent"></div>
+	</div>
+	
+	{#if showHelperText}
+		<p class="mt-2 sm:mt-3 text-center text-xs sm:text-xs text-exo-light-gray tracking-[0.1em] sm:tracking-[0.15em] uppercase">
+			<kbd class="px-1 sm:px-1.5 py-0.5 rounded bg-exo-medium-gray/30 text-exo-light-gray border border-exo-medium-gray/50">ENTER</kbd>
+			<span class="mx-0.5 sm:mx-1">TO SEND</span>
+			<span class="text-exo-medium-gray mx-1 sm:mx-2">|</span>
+			<kbd class="px-1 sm:px-1.5 py-0.5 rounded bg-exo-medium-gray/30 text-exo-light-gray border border-exo-medium-gray/50">SHIFT+ENTER</kbd>
+			<span class="mx-0.5 sm:mx-1">NEW LINE</span>
+			<span class="text-exo-medium-gray mx-1 sm:mx-2">|</span>
+			<span class="text-exo-light-gray">DRAG & DROP OR PASTE FILES</span>
+		</p>
+	{/if}
+</form>

dashboard/src/lib/components/ChatMessages.svelte

@@ -0,0 +1,462 @@
+<script lang="ts">
+	import { 
+		messages, 
+		currentResponse, 
+		isLoading,
+		deleteMessage,
+		editAndRegenerate,
+		regenerateLastResponse
+	} from '$lib/stores/app.svelte';
+	import type { MessageAttachment } from '$lib/stores/app.svelte';
+import { tick, onDestroy } from 'svelte';
+
+interface Props {
+	class?: string;
+	scrollParent?: HTMLElement | null;
+}
+
+let { class: className = '', scrollParent = null }: Props = $props();
+
+	const messageList = $derived(messages());
+	const response = $derived(currentResponse());
+	const loading = $derived(isLoading());
+
+// Ref for scroll anchor at bottom
+let scrollAnchorRef: HTMLDivElement | undefined = $state();
+
+// Scroll management
+const SCROLL_BOTTOM_THRESHOLD = 120;
+let autoScrollEnabled = true;
+let currentScrollEl: HTMLElement | null = null;
+
+function resolveScrollElement(): HTMLElement | null {
+	if (scrollParent) return scrollParent;
+	let node: HTMLElement | null = scrollAnchorRef?.parentElement as HTMLElement | null;
+	while (node) {
+		const isScrollable = node.scrollHeight > node.clientHeight + 1;
+		if (isScrollable) return node;
+		node = node.parentElement;
+	}
+	return null;
+}
+
+function handleScroll() {
+	if (!currentScrollEl) return;
+	const distanceFromBottom = currentScrollEl.scrollHeight - currentScrollEl.scrollTop - currentScrollEl.clientHeight;
+	const isNearBottom = distanceFromBottom < SCROLL_BOTTOM_THRESHOLD;
+	autoScrollEnabled = isNearBottom;
+}
+
+function attachScrollListener() {
+	const nextEl = resolveScrollElement();
+	if (currentScrollEl === nextEl) return;
+	if (currentScrollEl) {
+		currentScrollEl.removeEventListener('scroll', handleScroll);
+	}
+	currentScrollEl = nextEl;
+	if (currentScrollEl) {
+		currentScrollEl.addEventListener('scroll', handleScroll);
+		// Initialize state based on current position
+		handleScroll();
+	}
+}
+
+onDestroy(() => {
+	if (currentScrollEl) {
+		currentScrollEl.removeEventListener('scroll', handleScroll);
+	}
+});
+
+$effect(() => {
+	// Re-evaluate scroll container if prop changes or after mount
+	scrollParent;
+	attachScrollListener();
+});
+
+// Auto-scroll to bottom when messages change or response updates, but only if user is near bottom
+$effect(() => {
+	// Track these values to trigger effect
+	const _ = messageList.length;
+	const __ = response;
+	const ___ = loading;
+	
+	tick().then(() => {
+		const el = currentScrollEl ?? resolveScrollElement();
+		if (!el || !scrollAnchorRef) return;
+		const distanceFromBottom = el.scrollHeight - el.scrollTop - el.clientHeight;
+		const isNearBottom = distanceFromBottom < SCROLL_BOTTOM_THRESHOLD;
+		if (autoScrollEnabled || isNearBottom) {
+			scrollAnchorRef.scrollIntoView({ behavior: 'smooth', block: 'end' });
+			autoScrollEnabled = true;
+		}
+	});
+});
+
+	// Edit state
+	let editingMessageId = $state<string | null>(null);
+	let editContent = $state('');
+	let editTextareaRef: HTMLTextAreaElement | undefined = $state();
+
+	// Delete confirmation state
+	let deleteConfirmId = $state<string | null>(null);
+
+// Copied state for feedback
+let copiedMessageId = $state<string | null>(null);
+let expandedThinkingMessageIds = $state<Set<string>>(new Set());
+
+	function formatTimestamp(timestamp: number): string {
+		return new Date(timestamp).toLocaleTimeString('en-US', { 
+			hour12: false,
+			hour: '2-digit', 
+			minute: '2-digit',
+			second: '2-digit'
+		});
+	}
+
+	function getAttachmentIcon(attachment: MessageAttachment): string {
+		switch (attachment.type) {
+			case 'image': return '🖼';
+			case 'text': return '📄';
+			default: return '📎';
+		}
+	}
+
+	function truncateName(name: string, maxLen: number = 25): string {
+		if (name.length <= maxLen) return name;
+		const ext = name.slice(name.lastIndexOf('.'));
+		const base = name.slice(0, name.lastIndexOf('.'));
+		const available = maxLen - ext.length - 3;
+		return base.slice(0, available) + '...' + ext;
+	}
+
+	async function handleCopy(content: string, messageId: string) {
+		try {
+			await navigator.clipboard.writeText(content);
+			copiedMessageId = messageId;
+			setTimeout(() => {
+				copiedMessageId = null;
+			}, 2000);
+		} catch (error) {
+			console.error('Failed to copy:', error);
+		}
+	}
+
+function toggleThinkingVisibility(messageId: string) {
+	const next = new Set(expandedThinkingMessageIds);
+	if (next.has(messageId)) {
+		next.delete(messageId);
+	} else {
+		next.add(messageId);
+	}
+	expandedThinkingMessageIds = next;
+}
+
+function isThinkingExpanded(messageId: string): boolean {
+	return expandedThinkingMessageIds.has(messageId);
+}
+
+	function handleStartEdit(messageId: string, content: string) {
+		editingMessageId = messageId;
+		editContent = content;
+		setTimeout(() => {
+			if (editTextareaRef) {
+				editTextareaRef.focus();
+				editTextareaRef.setSelectionRange(editTextareaRef.value.length, editTextareaRef.value.length);
+				// Auto-resize
+				editTextareaRef.style.height = 'auto';
+				editTextareaRef.style.height = Math.min(editTextareaRef.scrollHeight, 200) + 'px';
+			}
+		}, 10);
+	}
+
+	function handleCancelEdit() {
+		editingMessageId = null;
+		editContent = '';
+	}
+
+	function handleSaveEdit() {
+		if (editingMessageId && editContent.trim()) {
+			editAndRegenerate(editingMessageId, editContent.trim());
+		}
+		editingMessageId = null;
+		editContent = '';
+	}
+
+	function handleEditKeydown(event: KeyboardEvent) {
+		if (event.key === 'Enter' && !event.shiftKey) {
+			event.preventDefault();
+			handleSaveEdit();
+		} else if (event.key === 'Escape') {
+			handleCancelEdit();
+		}
+	}
+
+	function handleEditInput() {
+		if (editTextareaRef) {
+			editTextareaRef.style.height = 'auto';
+			editTextareaRef.style.height = Math.min(editTextareaRef.scrollHeight, 200) + 'px';
+		}
+	}
+
+	function handleDeleteClick(messageId: string) {
+		deleteConfirmId = messageId;
+	}
+
+	function handleConfirmDelete() {
+		if (deleteConfirmId) {
+			deleteMessage(deleteConfirmId);
+			deleteConfirmId = null;
+		}
+	}
+
+	function handleCancelDelete() {
+		deleteConfirmId = null;
+	}
+
+	function handleRegenerate() {
+		regenerateLastResponse();
+	}
+
+	// Check if a message is the last assistant message
+	function isLastAssistantMessage(messageId: string): boolean {
+		for (let i = messageList.length - 1; i >= 0; i--) {
+			if (messageList[i].role === 'assistant') {
+				return messageList[i].id === messageId;
+			}
+		}
+		return false;
+	}
+</script>
+
+<div class="flex flex-col gap-4 sm:gap-6 {className}">
+	{#each messageList as message (message.id)}
+		<div class="group flex {message.role === 'user' ? 'justify-end' : 'justify-start'}">
+			<div class="{message.role === 'user' ? 'max-w-[85%] sm:max-w-[70%] flex flex-col items-end' : 'max-w-[95%] sm:max-w-[85%]'}">
+				{#if message.role === 'assistant'}
+					<!-- Assistant message header -->
+					<div class="flex items-center gap-1.5 sm:gap-2 mb-1.5 sm:mb-2">
+						<div class="w-1.5 h-1.5 sm:w-2 sm:h-2 bg-exo-yellow rounded-full shadow-[0_0_10px_rgba(255,215,0,0.5)]"></div>
+						<span class="text-sm sm:text-xs text-exo-yellow tracking-[0.15em] sm:tracking-[0.2em] uppercase font-medium">EXO</span>
+						<span class="text-xs sm:text-sm text-exo-light-gray tracking-wider tabular-nums">{formatTimestamp(message.timestamp)}</span>
+						{#if message.ttftMs || message.tps}
+							<span class="text-xs text-exo-light-gray/80 font-mono ml-2">
+								{#if message.ttftMs}<span class="text-exo-light-gray/50">TTFT</span> {message.ttftMs.toFixed(0)}ms{/if}{#if message.ttftMs && message.tps}<span class="text-exo-light-gray/30 mx-1">•</span>{/if}{#if message.tps}{message.tps.toFixed(1)} <span class="text-exo-light-gray/50">tok/s</span>{/if}
+							</span>
+						{/if}
+					</div>
+				{:else}
+					<!-- User message header -->
+					<div class="flex items-center justify-end gap-1.5 sm:gap-2 mb-1.5 sm:mb-2">
+						<span class="text-xs sm:text-sm text-exo-light-gray tracking-wider tabular-nums">{formatTimestamp(message.timestamp)}</span>
+						<span class="text-sm sm:text-xs text-exo-light-gray tracking-[0.1em] sm:tracking-[0.15em] uppercase">QUERY</span>
+						<div class="w-1.5 h-1.5 sm:w-2 sm:h-2 bg-exo-light-gray/50 rounded-full"></div>
+					</div>
+				{/if}
+				
+				{#if deleteConfirmId === message.id}
+					<!-- Delete confirmation -->
+					<div class="bg-red-500/10 border border-red-500/30 rounded-lg p-3">
+						<p class="text-xs text-red-400 mb-3">Delete this message{message.role === 'user' ? ' and all responses after it' : ''}?</p>
+						<div class="flex gap-2 justify-end">
+						<button
+							onclick={handleCancelDelete}
+							class="px-3 py-1.5 text-sm font-mono tracking-wider uppercase bg-exo-medium-gray/20 text-exo-light-gray border border-exo-medium-gray/30 rounded hover:bg-exo-medium-gray/30 transition-colors cursor-pointer"
+						>
+							CANCEL
+						</button>
+						<button
+							onclick={handleConfirmDelete}
+							class="px-3 py-1.5 text-sm font-mono tracking-wider uppercase bg-red-500/20 text-red-400 border border-red-500/30 rounded hover:bg-red-500/30 transition-colors cursor-pointer"
+						>
+								DELETE
+							</button>
+						</div>
+					</div>
+				{:else if editingMessageId === message.id}
+					<!-- Edit mode -->
+					<div class="command-panel rounded-lg p-3">
+						<textarea
+							bind:this={editTextareaRef}
+							bind:value={editContent}
+							onkeydown={handleEditKeydown}
+							oninput={handleEditInput}
+							class="w-full bg-exo-black/60 border border-exo-yellow/30 rounded px-3 py-2 text-sm text-foreground font-mono focus:outline-none focus:border-exo-yellow/50 resize-none"
+							style="min-height: 60px; max-height: 200px;"
+						></textarea>
+						<div class="flex gap-2 justify-end mt-2">
+						<button
+							onclick={handleCancelEdit}
+							class="px-3 py-1.5 text-sm font-mono tracking-wider uppercase bg-exo-medium-gray/20 text-exo-light-gray border border-exo-medium-gray/30 rounded hover:bg-exo-medium-gray/30 transition-colors cursor-pointer"
+						>
+							CANCEL
+						</button>
+						<button
+							onclick={handleSaveEdit}
+							disabled={!editContent.trim()}
+							class="px-3 py-1.5 text-sm font-mono tracking-wider uppercase bg-transparent text-exo-yellow border border-exo-yellow/30 rounded hover:border-exo-yellow/50 transition-colors disabled:opacity-50 disabled:cursor-not-allowed flex items-center gap-1.5 cursor-pointer"
+						>
+								<svg class="w-3 h-3" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M12 19l9 2-9-18-9 18 9-2zm0 0v-8" />
+								</svg>
+								SEND
+							</button>
+						</div>
+					</div>
+				{:else}
+					<div class="{message.role === 'user' 
+						? 'command-panel rounded-lg rounded-tr-sm inline-block' 
+						: 'command-panel rounded-lg rounded-tl-sm border-l-2 border-l-exo-yellow/50 inline-block'}">
+						
+						{#if message.role === 'user'}
+							<!-- User message styling -->
+							<div class="px-4 py-3">
+								<!-- Attachments -->
+								{#if message.attachments && message.attachments.length > 0}
+									<div class="flex flex-wrap gap-2 mb-3">
+										{#each message.attachments as attachment}
+											<div class="flex items-center gap-2 bg-exo-dark-gray/60 border border-exo-yellow/20 rounded px-2 py-1 text-xs font-mono">
+												{#if attachment.type === 'image' && attachment.preview}
+													<img 
+														src={attachment.preview} 
+														alt={attachment.name}
+														class="w-12 h-12 object-cover rounded border border-exo-yellow/20"
+													/>
+												{:else}
+													<span>{getAttachmentIcon(attachment)}</span>
+												{/if}
+												<span class="text-exo-yellow" title={attachment.name}>{truncateName(attachment.name)}</span>
+											</div>
+										{/each}
+									</div>
+								{/if}
+								
+								{#if message.content}
+									<div class="text-sm text-foreground font-mono tracking-wide whitespace-pre-wrap break-words leading-relaxed">
+										{message.content}
+									</div>
+								{/if}
+							</div>
+						{:else}
+							<!-- Assistant message styling -->
+							<div class="p-3 sm:p-4">
+								{#if message.thinking && message.thinking.trim().length > 0}
+									<div class="mb-3 rounded border border-exo-yellow/20 bg-exo-black/40">
+										<button
+											type="button"
+											class="w-full flex items-center justify-between px-3 py-2 text-xs font-mono uppercase tracking-[0.2em] text-exo-light-gray/80 hover:text-exo-yellow transition-colors cursor-pointer"
+											onclick={() => toggleThinkingVisibility(message.id)}
+											aria-expanded={isThinkingExpanded(message.id)}
+											aria-controls={`thinking-panel-${message.id}`}
+										>
+											<span class="flex items-center gap-2 tracking-[0.25em]">
+												<svg
+													class={`w-3.5 h-3.5 text-current transition-transform duration-200 ${isThinkingExpanded(message.id) ? 'rotate-90' : ''}`}
+													fill="none"
+													viewBox="0 0 24 24"
+													stroke="currentColor"
+													aria-hidden="true"
+												>
+													<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 5l7 7-7 7" />
+												</svg>
+												<span>Thinking...</span>
+											</span>
+											<span class="text-[10px] tracking-[0.2em] text-exo-light-gray/60">
+												{isThinkingExpanded(message.id) ? 'HIDE' : 'SHOW'}
+											</span>
+										</button>
+										{#if isThinkingExpanded(message.id)}
+											<div
+												id={`thinking-panel-${message.id}`}
+												class="px-3 pb-3 text-xs text-exo-light-gray/90 font-mono whitespace-pre-wrap break-words leading-relaxed"
+											>
+												{message.thinking.trim()}
+											</div>
+										{/if}
+									</div>
+								{/if}
+								<div class="text-sm text-foreground font-mono tracking-wide whitespace-pre-wrap break-words leading-relaxed">
+									{message.content || (loading ? response : '')}
+									{#if loading && !message.content}
+										<span class="inline-block w-2 h-4 bg-exo-yellow/70 ml-1 cursor-blink"></span>
+									{/if}
+								</div>
+							</div>
+						{/if}
+					</div>
+					
+					<!-- Action buttons -->
+					<div class="flex items-center gap-1 mt-1.5 opacity-0 group-hover:opacity-100 transition-opacity {message.role === 'user' ? 'justify-end' : 'justify-start'}">
+						<!-- Copy button -->
+						<button
+							onclick={() => handleCopy(message.content, message.id)}
+							class="p-1.5 text-exo-light-gray hover:text-exo-yellow transition-colors rounded cursor-pointer"
+							title="Copy message"
+						>
+							{#if copiedMessageId === message.id}
+								<svg class="w-3.5 h-3.5 text-green-400" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M5 13l4 4L19 7" />
+								</svg>
+							{:else}
+								<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z" />
+								</svg>
+							{/if}
+						</button>
+						
+						<!-- Edit button (user messages only) -->
+						{#if message.role === 'user'}
+							<button
+								onclick={() => handleStartEdit(message.id, message.content)}
+								class="p-1.5 text-exo-light-gray hover:text-exo-yellow transition-colors rounded cursor-pointer"
+								title="Edit message"
+							>
+								<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z" />
+								</svg>
+							</button>
+						{/if}
+						
+						<!-- Regenerate button (last assistant message only) -->
+						{#if message.role === 'assistant' && isLastAssistantMessage(message.id) && !loading}
+							<button
+								onclick={handleRegenerate}
+								class="p-1.5 text-exo-light-gray hover:text-exo-yellow transition-colors rounded cursor-pointer"
+								title="Regenerate response"
+							>
+								<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 4v5h.582m15.356 2A8.001 8.001 0 004.582 9m0 0H9m11 11v-5h-.581m0 0a8.003 8.003 0 01-15.357-2m15.357 2H15" />
+								</svg>
+							</button>
+						{/if}
+						
+						<!-- Delete button -->
+						<button
+							onclick={() => handleDeleteClick(message.id)}
+							class="p-1.5 text-exo-light-gray hover:text-red-400 transition-colors rounded hover:bg-red-500/10 cursor-pointer"
+							title="Delete message"
+						>
+							<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+								<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 7l-.867 12.142A2 2 0 0116.138 21H7.862a2 2 0 01-1.995-1.858L5 7m5 4v6m4-6v6m1-10V4a1 1 0 00-1-1h-4a1 1 0 00-1 1v3M4 7h16" />
+							</svg>
+						</button>
+					</div>
+				{/if}
+			</div>
+		</div>
+	{/each}
+	
+	{#if messageList.length === 0}
+		<div class="flex-1 flex flex-col items-center justify-center text-center pt-[20vh]">
+			<div class="w-12 h-12 sm:w-16 sm:h-16 border border-exo-yellow/20 rounded-full flex items-center justify-center mb-3 sm:mb-4">
+				<div class="w-6 h-6 sm:w-8 sm:h-8 border border-exo-yellow/40 rounded-full flex items-center justify-center">
+					<div class="w-1.5 h-1.5 sm:w-2 sm:h-2 bg-exo-yellow/60 rounded-full"></div>
+				</div>
+			</div>
+			<p class="text-xs sm:text-sm text-exo-light-gray tracking-[0.15em] sm:tracking-[0.2em] uppercase">AWAITING INPUT</p>
+			<p class="text-sm sm:text-xs text-exo-light-gray tracking-wider mt-1">ENTER A QUERY TO BEGIN</p>
+		</div>
+	{/if}
+	
+	<!-- Scroll anchor for auto-scroll -->
+	<div bind:this={scrollAnchorRef}></div>
+</div>

dashboard/src/lib/components/ChatSidebar.svelte

@@ -0,0 +1,430 @@
+<script lang="ts">
+import { 
+		conversations, 
+		activeConversationId, 
+		createConversation, 
+		loadConversation, 
+		deleteConversation, 
+		deleteAllConversations,
+		renameConversation,
+		clearChat,
+		instances,
+		debugMode,
+		toggleDebugMode
+	} from '$lib/stores/app.svelte';
+
+	interface Props {
+		class?: string;
+	}
+
+	let { class: className = '' }: Props = $props();
+
+	const conversationList = $derived(conversations());
+	const activeId = $derived(activeConversationId());
+const instanceData = $derived(instances());
+const debugEnabled = $derived(debugMode());
+
+	let searchQuery = $state('');
+	let editingId = $state<string | null>(null);
+	let editingName = $state('');
+	let deleteConfirmId = $state<string | null>(null);
+	let showDeleteAllConfirm = $state(false);
+
+	const filteredConversations = $derived(
+		searchQuery.trim()
+			? conversationList.filter(c => c.name.toLowerCase().includes(searchQuery.toLowerCase()))
+			: conversationList
+	);
+
+	function handleNewChat() {
+		createConversation();
+	}
+
+	function handleSelectConversation(id: string) {
+		loadConversation(id);
+	}
+
+	function handleStartEdit(id: string, name: string, event: MouseEvent) {
+		event.stopPropagation();
+		editingId = id;
+		editingName = name;
+	}
+
+	function handleSaveEdit() {
+		if (editingId && editingName.trim()) {
+			renameConversation(editingId, editingName.trim());
+		}
+		editingId = null;
+		editingName = '';
+	}
+
+	function handleCancelEdit() {
+		editingId = null;
+		editingName = '';
+	}
+
+	function handleEditKeydown(event: KeyboardEvent) {
+		if (event.key === 'Enter') {
+			handleSaveEdit();
+		} else if (event.key === 'Escape') {
+			handleCancelEdit();
+		}
+	}
+
+	function handleDeleteClick(id: string, event: MouseEvent) {
+		event.stopPropagation();
+		deleteConfirmId = id;
+	}
+
+	function handleConfirmDelete() {
+		if (deleteConfirmId) {
+			deleteConversation(deleteConfirmId);
+			deleteConfirmId = null;
+		}
+	}
+
+	function handleCancelDelete() {
+		deleteConfirmId = null;
+	}
+
+	function handleDeleteAllClick() {
+		showDeleteAllConfirm = true;
+	}
+
+	function handleConfirmDeleteAll() {
+		deleteAllConversations();
+		showDeleteAllConfirm = false;
+	}
+
+	function handleCancelDeleteAll() {
+		showDeleteAllConfirm = false;
+	}
+
+	function formatDate(timestamp: number): string {
+		const date = new Date(timestamp);
+		const now = new Date();
+		const diffDays = Math.floor((now.getTime() - date.getTime()) / (1000 * 60 * 60 * 24));
+		
+		if (diffDays === 0) {
+			return date.toLocaleTimeString('en-US', { hour: '2-digit', minute: '2-digit' });
+		} else if (diffDays === 1) {
+			return 'Yesterday';
+		} else if (diffDays < 7) {
+			return date.toLocaleDateString('en-US', { weekday: 'short' });
+		} else {
+			return date.toLocaleDateString('en-US', { month: 'short', day: 'numeric' });
+		}
+	}
+
+	function getLastAssistantStats(conversation: typeof conversationList[0]): { ttftMs?: number; tps?: number } | null {
+		// Find the last assistant message with stats
+		for (let i = conversation.messages.length - 1; i >= 0; i--) {
+			const msg = conversation.messages[i];
+			if (msg.role === 'assistant' && (msg.ttftMs || msg.tps)) {
+				return { ttftMs: msg.ttftMs, tps: msg.tps };
+			}
+		}
+		return null;
+	}
+
+	function formatModelName(modelId: string | null | undefined): string {
+		if (!modelId) return 'Unknown Model';
+		const parts = modelId.split('/');
+		const tail = parts[parts.length - 1] || modelId;
+		return tail || modelId;
+	}
+
+	function formatStrategy(sharding: string | null | undefined, instanceType: string | null | undefined): string {
+		const shardLabel = sharding ?? 'Unknown';
+		const typeLabel = instanceType ?? null;
+		return typeLabel ? `${shardLabel} (${typeLabel})` : shardLabel;
+	}
+
+	function getTaggedValue(obj: unknown): [string | null, unknown] {
+		if (!obj || typeof obj !== 'object') return [null, null];
+		const keys = Object.keys(obj as Record<string, unknown>);
+		if (keys.length === 1) {
+			return [keys[0], (obj as Record<string, unknown>)[keys[0]]];
+		}
+		return [null, null];
+	}
+
+	function extractInstanceModelId(instanceWrapped: unknown): string | null {
+		const [, instance] = getTaggedValue(instanceWrapped);
+		if (!instance || typeof instance !== 'object') return null;
+		const inst = instance as { shardAssignments?: { modelId?: string } };
+		return inst.shardAssignments?.modelId ?? null;
+	}
+
+	function describeInstance(instanceWrapped: unknown): { sharding: string | null; instanceType: string | null } {
+		const [instanceTag, instance] = getTaggedValue(instanceWrapped);
+		if (!instance || typeof instance !== 'object') {
+			return { sharding: null, instanceType: null };
+		}
+
+		let instanceType: string | null = null;
+		if (instanceTag === 'MlxRingInstance') instanceType = 'MLX Ring';
+		else if (instanceTag === 'MlxIbvInstance' || instanceTag === 'MlxJacclInstance') instanceType = 'MLX RDMA';
+
+		let sharding: string | null = null;
+		const inst = instance as { shardAssignments?: { runnerToShard?: Record<string, unknown> } };
+		const runnerToShard = inst.shardAssignments?.runnerToShard || {};
+		const firstShardWrapped = Object.values(runnerToShard)[0];
+		if (firstShardWrapped) {
+			const [shardTag] = getTaggedValue(firstShardWrapped);
+			if (shardTag === 'PipelineShardMetadata') sharding = 'Pipeline';
+			else if (shardTag === 'TensorShardMetadata') sharding = 'Tensor';
+			else if (shardTag === 'PrefillDecodeShardMetadata') sharding = 'Prefill/Decode';
+		}
+
+		return { sharding, instanceType };
+	}
+
+	function resolveConversationInfo(conversation: typeof conversationList[0]): { modelLabel: string; strategyLabel: string } {
+		// Attempt to match conversation model to an instance
+		let matchedInstance: unknown = null;
+		let modelId = conversation.modelId ?? null;
+
+		if (modelId) {
+			for (const [, instanceWrapper] of Object.entries(instanceData)) {
+				const candidate = extractInstanceModelId(instanceWrapper);
+				if (candidate === modelId) {
+					matchedInstance = instanceWrapper;
+					break;
+				}
+			}
+		}
+
+		// Fallback: use the first available instance if no explicit match
+		if (!matchedInstance) {
+			const firstInstance = Object.values(instanceData)[0];
+			if (firstInstance) {
+				matchedInstance = firstInstance;
+				modelId = modelId ?? extractInstanceModelId(firstInstance);
+			}
+		}
+
+		const instanceDetails = matchedInstance ? describeInstance(matchedInstance) : { sharding: null, instanceType: null };
+		const displayModel = modelId ?? conversation.modelId ?? null;
+		const sharding = conversation.sharding ?? instanceDetails.sharding ?? 'Unknown';
+		const instanceType = conversation.instanceType ?? instanceDetails.instanceType;
+
+		return {
+			modelLabel: formatModelName(displayModel),
+			strategyLabel: formatStrategy(sharding, instanceType)
+		};
+	}
+</script>
+
+<aside class="flex flex-col h-full bg-exo-dark-gray border-r border-exo-yellow/10 {className}">
+	<!-- Header -->
+	<div class="p-4">
+		<button
+			onclick={handleNewChat}
+			class="w-full flex items-center justify-center gap-2 py-2.5 px-4 bg-transparent border border-exo-yellow/30 text-exo-yellow text-xs font-mono tracking-wider uppercase hover:border-exo-yellow/50 transition-all cursor-pointer"
+		>
+			<svg class="w-4 h-4" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+				<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M12 4v16m8-8H4" />
+			</svg>
+			NEW CHAT
+		</button>
+	</div>
+
+	<!-- Search -->
+	<div class="px-4 py-3">
+		<div class="relative">
+			<svg class="absolute left-3 top-1/2 -translate-y-1/2 w-3.5 h-3.5 text-white/50" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+				<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M21 21l-6-6m2-5a7 7 0 11-14 0 7 7 0 0114 0z" />
+			</svg>
+			<input
+				type="text"
+				bind:value={searchQuery}
+				placeholder="Search conversations..."
+				class="w-full bg-exo-black/40 border border-exo-medium-gray/30 rounded px-3 py-2 pl-9 text-xs text-white/90 placeholder:text-white/40 focus:outline-none focus:border-exo-yellow/30"
+			/>
+		</div>
+	</div>
+
+	<!-- Conversation List -->
+	<div class="flex-1 overflow-y-auto">
+		{#if filteredConversations.length > 0}
+			<div class="py-2">
+				<div class="px-4 py-2">
+					<span class="text-sm text-white/70 font-mono tracking-wider uppercase">
+						{searchQuery ? 'SEARCH RESULTS' : 'CONVERSATIONS'}
+					</span>
+				</div>
+				
+				{#each filteredConversations as conversation (conversation.id)}
+					{@const info = resolveConversationInfo(conversation)}
+					<div class="px-2">
+						{#if editingId === conversation.id}
+							<!-- Edit mode -->
+							<div class="p-2 bg-transparent border border-exo-yellow/20 rounded mb-1">
+								<input
+									type="text"
+									bind:value={editingName}
+									onkeydown={handleEditKeydown}
+									class="w-full bg-exo-black/60 border border-exo-yellow/30 rounded px-2 py-1.5 text-xs text-exo-light-gray focus:outline-none focus:border-exo-yellow/50 mb-2"
+									autofocus
+								/>
+								<div class="flex gap-2">
+									<button
+										onclick={handleSaveEdit}
+										class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-transparent text-exo-yellow border border-exo-yellow/30 rounded hover:border-exo-yellow/50 cursor-pointer"
+									>
+										SAVE
+									</button>
+									<button
+										onclick={handleCancelEdit}
+										class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-medium-gray/20 text-exo-light-gray border border-exo-medium-gray/30 rounded hover:bg-exo-medium-gray/30 cursor-pointer"
+									>
+										CANCEL
+									</button>
+								</div>
+							</div>
+						{:else if deleteConfirmId === conversation.id}
+							<!-- Delete confirmation -->
+							<div class="p-2 bg-red-500/10 border border-red-500/30 rounded mb-1">
+								<p class="text-xs text-red-400 mb-2">Delete "{conversation.name}"?</p>
+								<div class="flex gap-2">
+									<button
+										onclick={handleConfirmDelete}
+										class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-red-500/20 text-red-400 border border-red-500/30 rounded hover:bg-red-500/30 cursor-pointer"
+									>
+										DELETE
+									</button>
+									<button
+										onclick={handleCancelDelete}
+										class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-medium-gray/20 text-exo-light-gray border border-exo-medium-gray/30 rounded hover:bg-exo-medium-gray/30 cursor-pointer"
+									>
+										CANCEL
+									</button>
+								</div>
+							</div>
+						{:else}
+							<!-- Normal view -->
+							{@const stats = getLastAssistantStats(conversation)}
+							<div
+								role="button"
+								tabindex="0"
+								onclick={() => handleSelectConversation(conversation.id)}
+								onkeydown={(e) => e.key === 'Enter' && handleSelectConversation(conversation.id)}
+								class="group w-full flex items-center justify-between p-2 rounded mb-1 transition-all text-left cursor-pointer
+									{activeId === conversation.id 
+								? 'bg-transparent border border-exo-yellow/30' 
+									: 'hover:border-exo-yellow/20 border border-transparent'}"
+							>
+								<div class="flex-1 min-w-0 pr-2">
+									<div class="text-sm truncate {activeId === conversation.id ? 'text-exo-yellow' : 'text-white/90'}">
+										{conversation.name}
+									</div>
+									<div class="text-sm text-white/50 mt-0.5">
+										{formatDate(conversation.updatedAt)}
+									</div>
+									<div class="text-sm text-white/70 truncate">
+										{info.modelLabel}
+									</div>
+									<div class="text-xs text-white/60 font-mono">
+										Strategy: <span class="text-white/80">{info.strategyLabel}</span>
+									</div>
+									{#if stats}
+										<div class="text-xs text-white/60 font-mono mt-1">
+											{#if stats.ttftMs}<span class="text-white/40">TTFT</span> {stats.ttftMs.toFixed(0)}ms{/if}{#if stats.ttftMs && stats.tps}<span class="text-white/30 mx-1.5">•</span>{/if}{#if stats.tps}{stats.tps.toFixed(1)} <span class="text-white/40">tok/s</span>{/if}
+										</div>
+									{/if}
+								</div>
+								
+								<div class="flex items-center gap-1 opacity-0 group-hover:opacity-100 transition-opacity">
+								<button
+									type="button"
+									onclick={(e) => handleStartEdit(conversation.id, conversation.name, e)}
+									class="p-1 text-exo-light-gray hover:text-exo-yellow transition-colors cursor-pointer"
+									title="Rename"
+								>
+										<svg class="w-3 h-3" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+											<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M11 5H6a2 2 0 00-2 2v11a2 2 0 002 2h11a2 2 0 002-2v-5m-1.414-9.414a2 2 0 112.828 2.828L11.828 15H9v-2.828l8.586-8.586z" />
+										</svg>
+									</button>
+								<button
+									type="button"
+									onclick={(e) => handleDeleteClick(conversation.id, e)}
+									class="p-1 text-exo-light-gray hover:text-red-400 transition-colors cursor-pointer"
+									title="Delete"
+								>
+										<svg class="w-3 h-3" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+											<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 7l-.867 12.142A2 2 0 0116.138 21H7.862a2 2 0 01-1.995-1.858L5 7m5 4v6m4-6v6m1-10V4a1 1 0 00-1-1h-4a1 1 0 00-1 1v3M4 7h16" />
+										</svg>
+									</button>
+								</div>
+							</div>
+						{/if}
+					</div>
+				{/each}
+			</div>
+		{:else}
+			<div class="flex flex-col items-center justify-center h-full p-4 text-center">
+				<div class="w-12 h-12 border border-exo-yellow/20 rounded-full flex items-center justify-center mb-3">
+					<svg class="w-6 h-6 text-exo-yellow/40" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+						<path stroke-linecap="round" stroke-linejoin="round" stroke-width="1.5" d="M8 12h.01M12 12h.01M16 12h.01M21 12c0 4.418-4.03 8-9 8a9.863 9.863 0 01-4.255-.949L3 20l1.395-3.72C3.512 15.042 3 13.574 3 12c0-4.418 4.03-8 9-8s9 3.582 9 8z" />
+					</svg>
+				</div>
+				<p class="text-xs text-white/70 font-mono tracking-wider uppercase mb-1">
+					{searchQuery ? 'NO RESULTS' : 'NO CONVERSATIONS'}
+				</p>
+				<p class="text-sm text-white/50">
+					{searchQuery ? 'Try a different search' : 'Start a new chat to begin'}
+				</p>
+			</div>
+		{/if}
+	</div>
+
+	<!-- Footer -->
+	<div class="p-3 border-t border-exo-yellow/10">
+		{#if showDeleteAllConfirm}
+			<div class="bg-red-500/10 border border-red-500/30 rounded p-2 mb-2">
+				<p class="text-xs text-red-400 text-center mb-2">Delete all {conversationList.length} conversations?</p>
+				<div class="flex gap-2">
+				<button
+					onclick={handleConfirmDeleteAll}
+					class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-red-500/20 text-red-400 border border-red-500/30 rounded hover:bg-red-500/30 transition-colors cursor-pointer"
+				>
+					DELETE ALL
+				</button>
+				<button
+					onclick={handleCancelDeleteAll}
+					class="flex-1 py-1.5 text-xs font-mono tracking-wider uppercase bg-exo-medium-gray/20 text-exo-light-gray border border-exo-medium-gray/30 rounded hover:bg-exo-medium-gray/30 transition-colors cursor-pointer"
+				>
+						CANCEL
+					</button>
+				</div>
+			</div>
+		{:else if conversationList.length > 0}
+		<button
+			onclick={handleDeleteAllClick}
+			class="w-full flex items-center justify-center gap-2 py-1.5 text-sm font-mono tracking-wider uppercase text-white/70 hover:text-red-400 hover:bg-red-500/10 border border-transparent hover:border-red-500/20 rounded transition-all cursor-pointer"
+		>
+				<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+					<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 7l-.867 12.142A2 2 0 0116.138 21H7.862a2 2 0 01-1.995-1.858L5 7m5 4v6m4-6v6m1-10V4a1 1 0 00-1-1h-4a1 1 0 00-1 1v3M4 7h16" />
+				</svg>
+				DELETE ALL CHATS
+			</button>
+		{/if}
+	<div class="flex items-center justify-center gap-3 {conversationList.length > 0 && !showDeleteAllConfirm ? 'mt-2' : ''}">
+		<button
+			type="button"
+			onclick={toggleDebugMode}
+			class="p-1.5 rounded border border-exo-medium-gray/40 hover:border-exo-yellow/50 transition-colors cursor-pointer"
+			title="Toggle debug mode"
+		>
+			<svg class="w-4 h-4 {debugEnabled ? 'text-exo-yellow' : 'text-exo-medium-gray'}" fill="currentColor" viewBox="0 0 24 24">
+				<path d="M19 8h-1.81A6.002 6.002 0 0 0 12 2a6.002 6.002 0 0 0-5.19 3H5a1 1 0 0 0 0 2h1v2H5a1 1 0 0 0 0 2h1v2H5a1 1 0 0 0 0 2h1.81A6.002 6.002 0 0 0 12 22a6.002 6.002 0 0 0 5.19-3H19a1 1 0 0 0 0-2h-1v-2h1a1 1 0 0 0 0-2h-1v-2h1a1 1 0 1 0 0-2Zm-5 10.32V19a1 1 0 1 1-2 0v-.68a3.999 3.999 0 0 1-3-3.83V9.32a3.999 3.999 0 0 1 3-3.83V5a1 1 0 0 1 2 0v.49a3.999 3.999 0 0 1 3 3.83v5.17a3.999 3.999 0 0 1-3 3.83Z"/>
+			</svg>
+		</button>
+		<div class="text-xs text-white/60 font-mono tracking-wider text-center">
+			{conversationList.length} CONVERSATION{conversationList.length !== 1 ? 'S' : ''}
+		</div>
+	</div>
+	</div>
+</aside>
+

dashboard/src/lib/components/HeaderNav.svelte

@@ -0,0 +1,57 @@
+<script lang="ts">
+	import { browser } from '$app/environment';
+
+	export let showHome = true;
+	export let onHome: (() => void) | null = null;
+
+	function handleHome(): void {
+		if (onHome) {
+			onHome();
+			return;
+		}
+		if (browser) {
+			// Hash router: send to root
+			window.location.hash = '/';
+		}
+	}
+</script>
+
+<header class="relative z-20 flex items-center justify-center px-6 pt-8 pb-4 bg-exo-dark-gray">
+	<!-- Center: Logo (clickable to go home) -->
+	<button
+		onclick={handleHome}
+		class="hover:opacity-80 transition-opacity {showHome ? 'cursor-pointer' : 'cursor-default'}"
+		title={showHome ? 'Go to home' : ''}
+		disabled={!showHome}
+	>
+		<img src="/exo-logo.png" alt="EXO" class="h-18 drop-shadow-[0_0_20px_rgba(255,215,0,0.5)]" />
+	</button>
+
+	<!-- Right: Home + Downloads -->
+	<div class="absolute right-6 top-1/2 -translate-y-1/2 flex items-center gap-4">
+		{#if showHome}
+			<button
+				onclick={handleHome}
+				class="text-sm text-exo-light-gray hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+				title="Back to topology view"
+			>
+				<svg class="w-4 h-4" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+					<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M3 12l2-2m0 0l7-7 7 7M5 10v10a1 1 0 001 1h3m10-11l2 2m-2-2v10a1 1 0 01-1 1h-3m-6 0a1 1 0 001-1v-4a1 1 0 011-1h2a1 1 0 011 1v4a1 1 0 001 1m-6 0h6" />
+				</svg>
+				Home
+			</button>
+		{/if}
+		<a
+			href="/#/downloads"
+			class="text-sm text-exo-light-gray hover:text-exo-yellow transition-colors tracking-wider uppercase flex items-center gap-2 cursor-pointer"
+			title="View downloads overview"
+		>
+			<svg class="w-4 h-4" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
+				<path d="M12 3v12" />
+				<path d="M7 12l5 5 5-5" />
+				<path d="M5 21h14" />
+			</svg>
+			Downloads
+		</a>
+	</div>
+</header>