parallelise Qwen3Next

## Motivation

Broken right now. We'll potentially add a better one later

## Changes

<!-- Describe what you changed in detail -->

## Why It Works

<!-- Explain why your approach solves the problem -->

## Test Plan

### Manual Testing
Used for evals without any issue.

### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->

bench/eval_config.toml

@@ -0,0 +1,66 @@
+# exo-eval configuration file
+# See bench/exo_eval.py for usage
+
+[eval]
+# Eval framework type: "lm_eval" | "swe_bench" | "custom"
+type = "lm_eval"
+# Require HuggingFace token (default: true)
+# Set to false if using only public datasets
+require_hf_token = true
+
+# Instance/placement configuration
+# Controls how exo sets up the model instance before running evals
+[instance]
+# Placement strategy: "ring" | "jaccl" | "both"
+instance_meta = "jaccl"
+# Sharding strategy: "pipeline" | "tensor" | "both"
+sharding = "tensor"
+# Node constraints
+min_nodes = 2
+max_nodes = 2
+
+# lm_eval configuration (EleutherAI's lm-evaluation-harness)
+[lm_eval]
+# Tasks to run (list of task names)
+# NOTE: Chat completions API only supports generation-based tasks.
+# Loglikelihood tasks (mmlu, hellaswag, arc) require /v1/completions endpoint.
+#
+# Generation-based tasks (work with chat completions):
+#   - mmlu_pro, mmlu_generative, mmlu_flan_cot_fewshot, mmlu_flan_cot_zeroshot
+#   - gsm8k, gsm8k_cot, gsm8k_cot_zeroshot
+#   - truthfulqa (uses generate_until for some subtasks)
+#   - humaneval, mbpp (code generation)
+#
+# Run `lm_eval --tasks list` to see all available tasks
+tasks = ["mmlu_pro"]
+# Number of few-shot examples (5 is standard for mmlu_pro CoT)
+num_fewshot = 5
+# Batch size (use 1 for API models, "auto" doesn't work)
+batch_size = 1
+# Number of concurrent requests (set > 1 to enable parallelism)
+# Higher values enable better batching throughput
+num_concurrent = 64
+# Apply chat template for instruct/chat models (default: true)
+apply_chat_template = true
+# Use fewshot examples as conversation turns (better for chat models)
+fewshot_as_multiturn = true
+# Optional: limit samples per task (omit or comment out for no limit)
+# limit = 100
+# Output path for results
+output_path = "bench/eval_results"
+
+# SWE-bench configuration (placeholder)
+[swe_bench]
+# SWE-bench dataset
+dataset = "princeton-nlp/SWE-bench_Lite"
+# Maximum workers for parallel execution
+max_workers = 8
+# Path for prediction outputs
+predictions_path = "bench/predictions"
+
+# Custom evaluation script configuration
+[custom]
+# Path to custom evaluation script
+script = "path/to/eval_script.py"
+# Arguments to pass to the script
+args = ["--arg1", "value1"]

bench/exo_eval.py

@@ -0,0 +1,679 @@
+#!/usr/bin/env python3
+# pyright: reportAny=false, reportUnknownMemberType=false, reportUnknownVariableType=false, reportUnknownArgumentType=false
+"""
+exo-eval: Evaluation harness for exo inference system.
+
+Supports multiple evaluation frameworks via TOML configuration:
+- lm_eval: Language model evaluation using EleutherAI's lm-evaluation-harness
+- swe_bench: SWE-bench evaluation (placeholder for future implementation)
+- custom: Custom evaluation scripts
+
+Usage:
+    uv run python -m bench.exo_eval --config bench/eval_config.toml --model Llama-3.2-1b-Instruct-4bit
+    uv run python -m bench.exo_eval --config bench/eval_config.toml --model Llama-3.2-1b-Instruct-4bit --dry-run
+"""
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import json
+import os
+import subprocess
+import sys
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Literal
+
+# Add parent directory to path for direct script execution
+if __name__ == "__main__" and __package__ is None:
+    sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+import tomlkit
+from huggingface_hub import get_token as get_hf_token
+from loguru import logger
+from tomlkit.exceptions import TOMLKitError
+
+from bench.exo_bench import (
+    ExoClient,
+    ExoHttpError,
+    instance_id_from_instance,
+    nodes_used_in_instance,
+    placement_filter,
+    resolve_model_short_id,
+    sharding_filter,
+    wait_for_instance_gone,
+    wait_for_instance_ready,
+)
+
+EvalType = Literal["lm_eval", "swe_bench", "custom"]
+
+
+def load_config(config_path: str) -> dict[str, Any]:
+    """Load and parse TOML configuration file."""
+    path = Path(config_path)
+    if not path.exists():
+        raise FileNotFoundError(f"Config file not found: {config_path}")
+
+    with open(path, encoding="utf-8") as f:
+        return dict(tomlkit.load(f))
+
+
+def get_eval_type(config: dict[str, Any]) -> EvalType:
+    """Extract evaluation type from config."""
+    eval_section = config.get("eval", {})
+    eval_type = eval_section.get("type", "lm_eval")
+    if eval_type not in ("lm_eval", "swe_bench", "custom"):
+        raise ValueError(f"Unknown eval type: {eval_type}")
+    return eval_type
+
+
+def check_hf_token(config: dict[str, Any]) -> bool:
+    """Check if HuggingFace token is available when required.
+
+    Returns True if token is available or not required, False otherwise.
+    """
+    eval_section = config.get("eval", {})
+    require_hf_token = eval_section.get("require_hf_token", True)
+
+    if not require_hf_token:
+        return True
+
+    token = get_hf_token()
+    if token is None:
+        logger.error(
+            "HuggingFace token not found. "
+            "Set HF_TOKEN environment variable or run 'huggingface-cli login'. "
+            "To disable this check, set require_hf_token = false in [eval] config."
+        )
+        return False
+
+    logger.info("HuggingFace token found")
+    return True
+
+
+def select_placement(
+    client: ExoClient,
+    full_model_id: str,
+    config: dict[str, Any],
+) -> dict[str, Any] | None:
+    """Select a placement based on config preferences."""
+    instance_config = config.get("instance", {})
+
+    # If explicit instance is provided, use it directly
+    if "instance" in instance_config:
+        return instance_config["instance"]
+
+    # Otherwise, select from previews based on preferences
+    instance_meta_pref = instance_config.get("instance_meta", "ring")
+    sharding_pref = instance_config.get("sharding", "pipeline")
+    max_nodes = instance_config.get("max_nodes", 4)
+    min_nodes = instance_config.get("min_nodes", 1)
+
+    previews_resp = client.request_json(
+        "GET", "/instance/previews", params={"model_id": full_model_id}
+    )
+    previews = previews_resp.get("previews") or []
+
+    selected: list[dict[str, Any]] = []
+    for p in previews:
+        if p.get("error") is not None:
+            continue
+        if not placement_filter(str(p.get("instance_meta", "")), instance_meta_pref):
+            continue
+        if not sharding_filter(str(p.get("sharding", "")), sharding_pref):
+            continue
+
+        instance = p.get("instance")
+        if not isinstance(instance, dict):
+            continue
+
+        n = nodes_used_in_instance(instance)
+        if min_nodes <= n <= max_nodes:
+            selected.append(p)
+
+    if not selected:
+        return None
+
+    # Sort by preference: exact match on sharding/meta, then by node count (descending)
+    def sort_key(p: dict[str, Any]) -> tuple[int, int, int]:
+        meta_match = (
+            1 if instance_meta_pref in str(p.get("instance_meta", "")).lower() else 0
+        )
+        sharding_match = 1 if sharding_pref in str(p.get("sharding", "")).lower() else 0
+        n_nodes = nodes_used_in_instance(p["instance"])
+        return (meta_match, sharding_match, n_nodes)
+
+    selected.sort(key=sort_key, reverse=True)
+    return selected[0]
+
+
+def setup_instance(
+    client: ExoClient,
+    full_model_id: str,
+    config: dict[str, Any],
+    dry_run: bool,
+) -> tuple[str | None, dict[str, Any] | None]:
+    """Create and wait for an instance to be ready. Returns (instance_id, preview)."""
+    preview = select_placement(client, full_model_id, config)
+
+    if preview is None:
+        logger.error("No valid placement found matching config preferences")
+        return None, None
+
+    instance_data = preview.get("instance")
+    instance: dict[str, Any] = (
+        instance_data if isinstance(instance_data, dict) else preview
+    )
+    instance_id = instance_id_from_instance(instance)
+
+    sharding = str(preview.get("sharding", "unknown"))
+    instance_meta = str(preview.get("instance_meta", "unknown"))
+    n_nodes = nodes_used_in_instance(instance)
+
+    logger.info(f"Selected placement: {sharding} / {instance_meta} / nodes={n_nodes}")
+    logger.info(f"Instance ID: {instance_id}")
+
+    if dry_run:
+        logger.info("[dry-run] Would create instance and wait for ready")
+        return instance_id, preview
+
+    # Create instance
+    client.request_json("POST", "/instance", body={"instance": instance})
+
+    try:
+        wait_for_instance_ready(client, instance_id)
+        logger.info("Instance is ready")
+        time.sleep(1)  # Brief pause after ready
+        return instance_id, preview
+    except (RuntimeError, TimeoutError) as e:
+        logger.error(f"Failed to initialize instance: {e}")
+        with contextlib.suppress(ExoHttpError):
+            client.request_json("DELETE", f"/instance/{instance_id}")
+        return None, None
+
+
+def teardown_instance(client: ExoClient, instance_id: str) -> None:
+    """Delete an instance and wait for it to be gone."""
+    try:
+        client.request_json("DELETE", f"/instance/{instance_id}")
+    except ExoHttpError as e:
+        if e.status != 404:
+            raise
+    except (ConnectionRefusedError, OSError):
+        logger.warning(
+            f"Could not connect to exo to delete instance {instance_id} (server may be down)"
+        )
+        return
+    try:
+        wait_for_instance_gone(client, instance_id)
+    except (ConnectionRefusedError, OSError, TimeoutError):
+        logger.warning("Could not verify instance deletion (server may be down)")
+        return
+    logger.info(f"Instance {instance_id} deleted")
+
+
+def build_lm_eval_args(
+    config: dict[str, Any],
+    base_url: str,
+    model: str,
+    output_path: str | None,
+    limit: int | None,
+    use_completions: bool,
+) -> list[str]:
+    """Build command-line arguments for lm_eval."""
+    lm_eval_config = config.get("lm_eval", {})
+
+    # Choose model type based on whether tasks need completions API
+    if use_completions:
+        model_type = "local-completions"
+        endpoint_url = f"{base_url}/v1/completions"
+    else:
+        model_type = "local-chat-completions"
+        endpoint_url = f"{base_url}/v1/chat/completions"
+
+    # Build model_args string with num_concurrent and timeout
+    model_args_parts = [f"model={model}", f"base_url={endpoint_url}"]
+    num_concurrent = lm_eval_config.get("num_concurrent")
+    if num_concurrent is not None and num_concurrent > 1:
+        model_args_parts.append(f"num_concurrent={num_concurrent}")
+    # Use a very long timeout (1 week) to handle large request queues
+    timeout = lm_eval_config.get("timeout", 604800)
+    model_args_parts.append(f"timeout={timeout}")
+    model_args = ",".join(model_args_parts)
+
+    args = [
+        sys.executable,
+        "-m",
+        "bench.lm_eval_patched",
+        "--model",
+        model_type,
+        "--model_args",
+        model_args,
+        "--verbosity",
+        "WARNING",
+    ]
+
+    # Tasks
+    tasks = lm_eval_config.get("tasks", ["mmlu"])
+    tasks_str = ",".join(tasks) if isinstance(tasks, list) else str(tasks)
+    args.extend(["--tasks", tasks_str])
+
+    # Few-shot
+    num_fewshot = lm_eval_config.get("num_fewshot")
+    if num_fewshot is not None:
+        args.extend(["--num_fewshot", str(num_fewshot)])
+
+    # Batch size (default to 1 for API models, "auto" doesn't work)
+    batch_size = lm_eval_config.get("batch_size", 1)
+    args.extend(["--batch_size", str(batch_size)])
+
+    # Apply chat template for instruct/chat models (default: true)
+    # Only applies to chat completions, but doesn't hurt to include
+    apply_chat_template = lm_eval_config.get("apply_chat_template", True)
+    if apply_chat_template and not use_completions:
+        args.append("--apply_chat_template")
+
+    # Fewshot as multiturn (optional, works with chat template)
+    fewshot_as_multiturn = lm_eval_config.get("fewshot_as_multiturn", False)
+    if fewshot_as_multiturn and not use_completions:
+        args.append("--fewshot_as_multiturn")
+
+    # Limit (command line overrides config)
+    effective_limit = limit if limit is not None else lm_eval_config.get("limit")
+    if effective_limit is not None:
+        args.extend(["--limit", str(effective_limit)])
+
+    # Output path
+    effective_output = output_path or lm_eval_config.get("output_path")
+    if effective_output:
+        args.extend(["--output_path", effective_output])
+        # Log model responses for post-hoc analysis when output is saved
+        args.append("--log_samples")
+
+    return args
+
+
+def run_lm_eval(
+    config: dict[str, Any],
+    host: str,
+    port: int,
+    model: str,
+    output_path: str | None,
+    limit: int | None,
+    dry_run: bool,
+) -> int:
+    """Run lm_eval evaluation."""
+    lm_eval_config = config.get("lm_eval", {})
+    tasks = lm_eval_config.get("tasks", ["mmlu"])
+    if isinstance(tasks, str):
+        tasks = [tasks]
+
+    exo_base_url = f"http://{host}:{port}"
+
+    # Build args - use native completions or chat completions endpoint directly
+    args = build_lm_eval_args(
+        config, exo_base_url, model, output_path, limit, use_completions=False
+    )
+    logger.info(f"lm_eval command: {' '.join(args)}")
+
+    if dry_run:
+        logger.info("[dry-run] Would execute the above command")
+        return 0
+
+    try:
+        result = subprocess.run(args, check=False)
+
+        # Print token usage summary from exo
+        try:
+            import httpx
+
+            usage_resp = httpx.get(f"{exo_base_url}/v1/usage", timeout=5)
+            if usage_resp.status_code == 200:
+                usage = usage_resp.json()
+                logger.info("--- Token Usage (Total) ---")
+                logger.info(f"  Requests:          {usage.get('total_requests', 0)}")
+                logger.info(
+                    f"  Prompt tokens:     {usage.get('total_prompt_tokens', 0)}"
+                )
+                logger.info(
+                    f"  Completion tokens: {usage.get('total_completion_tokens', 0)}"
+                )
+                logger.info(
+                    f"  Reasoning tokens:  {usage.get('total_reasoning_tokens', 0)}"
+                )
+                logger.info(f"  Total tokens:      {usage.get('total_tokens', 0)}")
+                by_model = usage.get("by_model", {})
+                if by_model:
+                    for model_name, counters in by_model.items():
+                        logger.info(f"--- Token Usage ({model_name}) ---")
+                        logger.info(
+                            f"  Requests:          {counters.get('requests', 0)}"
+                        )
+                        logger.info(
+                            f"  Prompt tokens:     {counters.get('prompt_tokens', 0)}"
+                        )
+                        logger.info(
+                            f"  Completion tokens: {counters.get('completion_tokens', 0)}"
+                        )
+                        logger.info(
+                            f"  Reasoning tokens:  {counters.get('reasoning_tokens', 0)}"
+                        )
+        except Exception:
+            pass  # Usage endpoint not available
+
+        return result.returncode
+    except FileNotFoundError:
+        logger.error("lm_eval not found. Install with: uv sync --extra eval")
+        return 1
+
+
+def run_swe_bench(
+    config: dict[str, Any],
+    host: str,
+    port: int,
+    model: str,
+    output_path: str | None,
+    dry_run: bool,
+) -> int:
+    """Run SWE-bench evaluation (placeholder)."""
+    swe_config = config.get("swe_bench", {})
+
+    dataset = swe_config.get("dataset", "princeton-nlp/SWE-bench_Lite")
+    max_workers = swe_config.get("max_workers", 8)
+    predictions_path = output_path or swe_config.get(
+        "predictions_path", "bench/predictions"
+    )
+
+    logger.info("SWE-bench evaluation configuration:")
+    logger.info(f"  Dataset: {dataset}")
+    logger.info(f"  Model: {model}")
+    logger.info(f"  API endpoint: http://{host}:{port}/v1")
+    logger.info(f"  Max workers: {max_workers}")
+    logger.info(f"  Predictions path: {predictions_path}")
+
+    if dry_run:
+        logger.info("[dry-run] SWE-bench evaluation would be executed")
+        return 0
+
+    logger.warning(
+        "SWE-bench integration is a placeholder. "
+        "Implement swebench inference and evaluation logic as needed."
+    )
+    return 0
+
+
+def run_custom_eval(
+    config: dict[str, Any],
+    host: str,
+    port: int,
+    model: str,
+    output_path: str | None,
+    dry_run: bool,
+) -> int:
+    """Run custom evaluation script."""
+    custom_config = config.get("custom", {})
+
+    script = custom_config.get("script")
+    if not script:
+        logger.error("No script specified in [custom] config section")
+        return 1
+
+    script_path = Path(script)
+    if not script_path.exists():
+        logger.error(f"Custom script not found: {script}")
+        return 1
+
+    script_args = custom_config.get("args", [])
+    if not isinstance(script_args, list):
+        script_args = [str(script_args)]
+
+    # Build environment with exo connection info
+    env = os.environ.copy()
+    env["EXO_HOST"] = host
+    env["EXO_PORT"] = str(port)
+    env["EXO_MODEL"] = model
+    if output_path:
+        env["EXO_OUTPUT_PATH"] = output_path
+
+    cmd = [sys.executable, str(script_path), *script_args]
+    logger.info(f"Custom eval command: {' '.join(cmd)}")
+
+    if dry_run:
+        logger.info("[dry-run] Would execute the above command")
+        return 0
+
+    result = subprocess.run(cmd, env=env, check=False)
+    return result.returncode
+
+
+def write_results_metadata(
+    output_path: str,
+    config: dict[str, Any],
+    host: str,
+    port: int,
+    model: str,
+    eval_type: EvalType,
+    return_code: int,
+    preview: dict[str, Any] | None,
+) -> None:
+    """Write evaluation metadata to a JSON file."""
+    metadata: dict[str, Any] = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "eval_type": eval_type,
+        "model": model,
+        "api_endpoint": f"http://{host}:{port}/v1",
+        "config": config,
+        "return_code": return_code,
+    }
+
+    if preview:
+        metadata["placement"] = {
+            "sharding": preview.get("sharding"),
+            "instance_meta": preview.get("instance_meta"),
+            "instance_id": instance_id_from_instance(preview["instance"])
+            if "instance" in preview
+            else None,
+        }
+
+    output_dir = Path(output_path)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    metadata_path = output_dir / "eval_metadata.json"
+
+    with open(metadata_path, "w", encoding="utf-8") as f:
+        json.dump(metadata, f, indent=2, ensure_ascii=False, default=str)
+
+    logger.info(f"Wrote evaluation metadata to: {metadata_path}")
+
+
+def main() -> int:
+    """Main entry point for exo-eval."""
+    ap = argparse.ArgumentParser(
+        prog="exo-eval",
+        description="Evaluation harness for exo inference system.",
+    )
+    ap.add_argument(
+        "--config",
+        required=True,
+        help="Path to TOML configuration file",
+    )
+    ap.add_argument(
+        "--host",
+        default=os.environ.get("EXO_HOST", "localhost"),
+        help="exo API host (default: localhost or EXO_HOST env var)",
+    )
+    ap.add_argument(
+        "--port",
+        type=int,
+        default=int(os.environ.get("EXO_PORT", "52415")),
+        help="exo API port (default: 52415 or EXO_PORT env var)",
+    )
+    ap.add_argument(
+        "--model",
+        required=True,
+        help="Model name/ID to evaluate",
+    )
+    ap.add_argument(
+        "--output",
+        default=None,
+        help="Output path for results (overrides config)",
+    )
+    ap.add_argument(
+        "--limit",
+        type=int,
+        default=None,
+        help="Limit samples per task (overrides config, lm_eval only)",
+    )
+    ap.add_argument(
+        "--timeout",
+        type=float,
+        default=604800.0,
+        help="HTTP timeout in seconds (default: 604800 = 1 week)",
+    )
+    ap.add_argument(
+        "--skip-instance-setup",
+        action="store_true",
+        help="Skip instance creation (assume instance already running)",
+    )
+    ap.add_argument(
+        "--pipeline",
+        type=int,
+        default=None,
+        metavar="N",
+        help="Use pipeline sharding with exactly N nodes (overrides config)",
+    )
+    ap.add_argument(
+        "--instance-meta",
+        choices=["ring", "jaccl", "both"],
+        default=None,
+        help="Instance meta preference (overrides config)",
+    )
+    ap.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Print commands without executing",
+    )
+    args = ap.parse_args()
+
+    logger.info(f"exo-eval starting with config: {args.config}")
+
+    try:
+        config = load_config(args.config)
+    except FileNotFoundError as e:
+        logger.error(str(e))
+        return 1
+    except TOMLKitError as e:
+        logger.error(f"Failed to parse config: {e}")
+        return 1
+
+    eval_type = get_eval_type(config)
+    logger.info(f"Evaluation type: {eval_type}")
+    logger.info(f"Model: {args.model}")
+    logger.info(f"API endpoint: http://{args.host}:{args.port}/v1")
+
+    # Apply CLI overrides to instance config
+    if args.pipeline is not None or args.instance_meta is not None:
+        instance_config = config.setdefault("instance", {})
+        if args.pipeline is not None:
+            instance_config["sharding"] = "pipeline"
+            instance_config["min_nodes"] = args.pipeline
+            instance_config["max_nodes"] = args.pipeline
+            logger.info(f"CLI override: pipeline={args.pipeline} nodes")
+            # Limit concurrency for pipeline to avoid GPU timeouts
+            if args.pipeline >= 2:
+                lm_eval_config = config.setdefault("lm_eval", {})
+                lm_eval_config["num_concurrent"] = 4
+                logger.info("CLI override: num_concurrent=4 (pipeline>=2)")
+        if args.instance_meta is not None:
+            instance_config["instance_meta"] = args.instance_meta
+            logger.info(f"CLI override: instance_meta={args.instance_meta}")
+
+    # Check HuggingFace token if required
+    if not check_hf_token(config):
+        return 1
+
+    # Setup instance and resolve model
+    instance_id: str | None = None
+    preview: dict[str, Any] | None = None
+    client: ExoClient | None = None
+
+    if args.skip_instance_setup:
+        # Use model name as-is when skipping instance setup
+        full_model_id = args.model
+        logger.info(f"Using model: {full_model_id} (instance setup skipped)")
+    else:
+        client = ExoClient(args.host, args.port, timeout_s=args.timeout)
+
+        # Resolve model
+        try:
+            short_id, full_model_id = resolve_model_short_id(client, args.model)
+            logger.info(f"Resolved model: {short_id} -> {full_model_id}")
+        except Exception as e:
+            logger.error(f"Failed to resolve model: {e}")
+            return 1
+
+        instance_id, preview = setup_instance(
+            client, full_model_id, config, args.dry_run
+        )
+        if instance_id is None and not args.dry_run:
+            return 1
+
+    try:
+        # Run evaluation
+        if eval_type == "lm_eval":
+            return_code = run_lm_eval(
+                config,
+                args.host,
+                args.port,
+                full_model_id,
+                args.output,
+                args.limit,
+                args.dry_run,
+            )
+        elif eval_type == "swe_bench":
+            return_code = run_swe_bench(
+                config,
+                args.host,
+                args.port,
+                full_model_id,
+                args.output,
+                args.dry_run,
+            )
+        elif eval_type == "custom":
+            return_code = run_custom_eval(
+                config,
+                args.host,
+                args.port,
+                full_model_id,
+                args.output,
+                args.dry_run,
+            )
+        else:
+            logger.error(f"Unknown eval type: {eval_type}")
+            return 1
+
+        # Write metadata if output path specified and not dry-run
+        output_path = args.output or config.get(eval_type, {}).get("output_path")
+        if output_path and not args.dry_run:
+            write_results_metadata(
+                output_path,
+                config,
+                args.host,
+                args.port,
+                full_model_id,
+                eval_type,
+                return_code,
+                preview,
+            )
+
+        return return_code
+
+    finally:
+        # Teardown instance
+        if instance_id and client and not args.skip_instance_setup and not args.dry_run:
+            teardown_instance(client, instance_id)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

bench/lm_eval_patched.py

@@ -0,0 +1,145 @@
+"""Patched lm_eval runner that fixes bugs in the upstream library.
+
+Fixes:
+- UnboundLocalError on `outputs` in TemplateAPI.amodel_call when API returns error
+- Prevents eval crash on transient API failures (returns None instead of raising)
+- Compatibility with transformers 5.x (missing AutoModelForVision2Seq)
+- sock_read timeout causing connection drops with large request queues
+
+Usage: python -m bench.lm_eval_patched [lm_eval args...]
+"""
+
+# ruff: noqa: I001, E402
+# pyright: reportMissingTypeStubs=false, reportUnknownVariableType=false
+# pyright: reportUnknownMemberType=false, reportAny=false, reportUnknownArgumentType=false
+# pyright: reportPrivateUsage=false, reportUnknownLambdaType=false
+
+# MUST patch transformers BEFORE any lm_eval imports
+# AutoModelForVision2Seq/AutoModelForImageTextToText were removed in transformers 5.0
+# Patch the lazy module's __getattr__ to return stubs for missing classes
+from transformers.utils import import_utils
+
+_original_getattr = import_utils._LazyModule.__getattr__
+
+
+def _patched_getattr(self: object, name: str) -> object:
+    if name in ("AutoModelForVision2Seq", "AutoModelForImageTextToText"):
+        return type(name, (), {})  # Return a stub class
+    return _original_getattr(self, name)  # type: ignore
+
+
+import_utils._LazyModule.__getattr__ = _patched_getattr
+
+import functools
+from typing import Any
+
+
+def _patch_amodel_call() -> None:
+    """Monkey-patch TemplateAPI.amodel_call to handle the unbound `outputs` variable bug."""
+    from lm_eval.models.api_models import TemplateAPI
+
+    original: Any = TemplateAPI.amodel_call
+
+    @functools.wraps(original)
+    async def patched_amodel_call(self: Any, *args: Any, **kwargs: Any) -> Any:
+        try:
+            return await original(self, *args, **kwargs)
+        except (UnboundLocalError, Exception):
+            # Return one empty-string result per request in the batch so the
+            # reorderer doesn't assert on missing coverage.
+            messages = kwargs.get("messages") or (args[2] if len(args) > 2 else [])
+            return [""] * max(len(messages), 1)
+
+    TemplateAPI.amodel_call = patched_amodel_call
+
+
+def _patch_client_timeout() -> None:
+    """Patch TemplateAPI.get_batched_requests to disable sock_read timeout.
+
+    By default, aiohttp's ClientTimeout can have a sock_read timeout that causes
+    connections to drop if no data is received for a while. With large request
+    queues, requests may wait a long time before processing starts, causing
+    spurious connection drops and retries that pile up requests.
+    """
+    from aiohttp import ClientSession, ClientTimeout, TCPConnector
+
+    from lm_eval.models.api_models import TemplateAPI
+
+    original_get_batched: Any = TemplateAPI.get_batched_requests
+
+    @functools.wraps(original_get_batched)
+    async def patched_get_batched_requests(self: Any, *args: Any, **kwargs: Any) -> Any:
+        # Override the timeout to explicitly disable sock_read timeout
+        # This prevents connection drops when requests are queued for a long time
+        original_timeout = getattr(self, "timeout", 604800)
+        conn = TCPConnector(limit=self._concurrent, ssl=self.verify_certificate)
+        timeout = ClientTimeout(
+            total=original_timeout, sock_read=None, sock_connect=None
+        )
+
+        async with ClientSession(connector=conn, timeout=timeout) as session:
+            # Call the internal async logic with our session
+            return await _run_batched_requests_with_session(
+                self, session, *args, **kwargs
+            )
+
+    async def _run_batched_requests_with_session(
+        self: Any,
+        session: ClientSession,
+        requests: Any,
+        cache_keys: Any = None,
+        ctxlens: Any = None,
+        **kwargs: Any,
+    ) -> Any:
+        import asyncio
+        import copy
+        import logging
+
+        from tqdm.asyncio import tqdm_asyncio
+        from tenacity import retry, stop_after_attempt, wait_exponential
+        from lm_eval.models.utils import chunks
+
+        eval_logger = logging.getLogger("lm_eval.models.api_models")
+        ctxlens = ctxlens if ctxlens else [None] * len(requests)
+        sem = asyncio.Semaphore(self._concurrent)
+
+        retry_: Any = retry(
+            stop=stop_after_attempt(self.max_retries),
+            wait=wait_exponential(multiplier=0.5, min=1, max=10),
+            reraise=True,
+            before_sleep=lambda retry_state: eval_logger.info(
+                f"Retry attempt {retry_state.attempt_number}"
+            ),
+        )(self.amodel_call)
+
+        tasks = [
+            asyncio.create_task(
+                retry_(
+                    session=session,
+                    sem=sem,
+                    messages=message,
+                    cache_keys=cache_key,
+                    ctxlens=ctxlen,
+                    gen_kwargs=copy.deepcopy(kwargs.get("gen_kwargs")),
+                    **{k: v for k, v in kwargs.items() if k != "gen_kwargs"},
+                )
+            )
+            for message, cache_key, ctxlen in zip(
+                chunks(requests, n=self._batch_size),
+                chunks(cache_keys, n=self._batch_size),
+                chunks(ctxlens, n=self._batch_size),
+                strict=True,
+            )
+        ]
+
+        return await tqdm_asyncio.gather(*tasks, desc="Requesting API")
+
+    TemplateAPI.get_batched_requests = patched_get_batched_requests
+
+
+if __name__ == "__main__":
+    _patch_amodel_call()
+    _patch_client_timeout()
+    from lm_eval.__main__ import cli_evaluate
+
+    cli_evaluate()

bench/stats_dashboard.html

@@ -0,0 +1,290 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>exo Usage Stats</title>
+<style>
+  * { margin: 0; padding: 0; box-sizing: border-box; }
+  body {
+    font-family: -apple-system, BlinkMacSystemFont, 'SF Mono', 'Menlo', monospace;
+    background: #1a1a2e;
+    color: #e0e0e0;
+    padding: 24px;
+    min-height: 100vh;
+  }
+  .header {
+    display: flex;
+    justify-content: space-between;
+    align-items: center;
+    margin-bottom: 24px;
+    padding-bottom: 16px;
+    border-bottom: 1px solid #333;
+  }
+  .header h1 {
+    font-size: 20px;
+    font-weight: 600;
+    color: #fff;
+  }
+  .status {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+    font-size: 13px;
+    color: #888;
+  }
+  .status-dot {
+    width: 8px;
+    height: 8px;
+    border-radius: 50%;
+    background: #666;
+  }
+  .status-dot.connected { background: #4caf50; }
+  .status-dot.error { background: #f44336; }
+  .config {
+    margin-bottom: 24px;
+    display: flex;
+    align-items: center;
+    gap: 8px;
+  }
+  .config label {
+    font-size: 12px;
+    color: #888;
+  }
+  .config input {
+    background: #252540;
+    border: 1px solid #444;
+    border-radius: 4px;
+    color: #e0e0e0;
+    padding: 4px 8px;
+    font-size: 13px;
+    font-family: inherit;
+    width: 280px;
+  }
+  .section {
+    background: #252540;
+    border-radius: 8px;
+    padding: 20px;
+    margin-bottom: 16px;
+  }
+  .section h2 {
+    font-size: 14px;
+    font-weight: 600;
+    color: #aaa;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+    margin-bottom: 16px;
+  }
+  .stat-grid {
+    display: grid;
+    grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+    gap: 16px;
+  }
+  .stat-card {
+    background: #1a1a2e;
+    border-radius: 6px;
+    padding: 16px;
+  }
+  .stat-label {
+    font-size: 11px;
+    color: #888;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+    margin-bottom: 4px;
+  }
+  .stat-value {
+    font-size: 28px;
+    font-weight: 700;
+    color: #fff;
+  }
+  .stat-rate {
+    font-size: 12px;
+    color: #4caf50;
+    margin-top: 4px;
+  }
+  table {
+    width: 100%;
+    border-collapse: collapse;
+    font-size: 13px;
+  }
+  th {
+    text-align: left;
+    padding: 8px 12px;
+    color: #888;
+    font-weight: 500;
+    border-bottom: 1px solid #333;
+    font-size: 11px;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+  }
+  td {
+    padding: 8px 12px;
+    border-bottom: 1px solid #2a2a45;
+  }
+  td.num {
+    text-align: right;
+    font-variant-numeric: tabular-nums;
+  }
+  .model-name {
+    color: #7c9eff;
+    max-width: 300px;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
+  }
+  .empty-state {
+    color: #666;
+    font-style: italic;
+    padding: 16px 0;
+  }
+</style>
+</head>
+<body>
+  <div class="header">
+    <h1>exo Usage Stats</h1>
+    <div class="status">
+      <div class="status-dot" id="statusDot"></div>
+      <span id="statusText">connecting...</span>
+    </div>
+  </div>
+
+  <div class="config">
+    <label for="baseUrl">Base URL:</label>
+    <input type="text" id="baseUrl" value="http://mac8-1:52415">
+  </div>
+
+  <div class="section">
+    <h2>Totals</h2>
+    <div class="stat-grid">
+      <div class="stat-card">
+        <div class="stat-label">Requests</div>
+        <div class="stat-value" id="totalRequests">0</div>
+      </div>
+      <div class="stat-card">
+        <div class="stat-label">Prompt Tokens</div>
+        <div class="stat-value" id="totalPrompt">0</div>
+        <div class="stat-rate" id="promptRate"></div>
+      </div>
+      <div class="stat-card">
+        <div class="stat-label">Completion Tokens</div>
+        <div class="stat-value" id="totalCompletion">0</div>
+        <div class="stat-rate" id="completionRate"></div>
+      </div>
+      <div class="stat-card">
+        <div class="stat-label">Reasoning Tokens</div>
+        <div class="stat-value" id="totalReasoning">0</div>
+      </div>
+      <div class="stat-card">
+        <div class="stat-label">Total Tokens</div>
+        <div class="stat-value" id="totalTokens">0</div>
+        <div class="stat-rate" id="totalRate"></div>
+      </div>
+    </div>
+  </div>
+
+  <div class="section">
+    <h2>Per-Model Breakdown</h2>
+    <div id="modelTable">
+      <div class="empty-state">No data yet</div>
+    </div>
+  </div>
+
+<script>
+
+  function fmt(n) {
+    return n.toLocaleString();
+  }
+
+  // Track first non-zero timestamp for overall average rate
+  let firstSeenTime = null;
+  let firstSeenTokens = { prompt: 0, completion: 0, total: 0 };
+
+  function setRate(id, currentTokens, tokenType) {
+    const el = document.getElementById(id);
+    if (firstSeenTime === null || currentTokens <= firstSeenTokens[tokenType]) {
+      el.textContent = '';
+      return;
+    }
+    const elapsed = (performance.now() / 1000) - firstSeenTime;
+    if (elapsed <= 0) { el.textContent = ''; return; }
+    const delta = currentTokens - firstSeenTokens[tokenType];
+    const avg = delta / elapsed;
+    el.textContent = fmt(Math.round(avg)) + ' tok/s avg';
+  }
+
+  function renderModelTable(byModel) {
+    const container = document.getElementById('modelTable');
+    const models = Object.entries(byModel);
+    if (models.length === 0) {
+      container.innerHTML = '<div class="empty-state">No data yet</div>';
+      return;
+    }
+    let html = '<table><thead><tr>';
+    html += '<th>Model</th><th style="text-align:right">Requests</th>';
+    html += '<th style="text-align:right">Prompt</th>';
+    html += '<th style="text-align:right">Completion</th>';
+    html += '<th style="text-align:right">Reasoning</th>';
+    html += '<th style="text-align:right">Total</th>';
+    html += '</tr></thead><tbody>';
+    for (const [name, counters] of models) {
+      const total = (counters.prompt_tokens || 0) + (counters.completion_tokens || 0);
+      html += '<tr>';
+      html += `<td class="model-name" title="${name}">${name}</td>`;
+      html += `<td class="num">${fmt(counters.requests || 0)}</td>`;
+      html += `<td class="num">${fmt(counters.prompt_tokens || 0)}</td>`;
+      html += `<td class="num">${fmt(counters.completion_tokens || 0)}</td>`;
+      html += `<td class="num">${fmt(counters.reasoning_tokens || 0)}</td>`;
+      html += `<td class="num">${fmt(total)}</td>`;
+      html += '</tr>';
+    }
+    html += '</tbody></table>';
+    container.innerHTML = html;
+  }
+
+  async function poll() {
+    const baseUrl = document.getElementById('baseUrl').value.replace(/\/+$/, '');
+    const dot = document.getElementById('statusDot');
+    const text = document.getElementById('statusText');
+
+    try {
+      const resp = await fetch(baseUrl + '/v1/usage');
+      if (!resp.ok) throw new Error(`HTTP ${resp.status}`);
+      const data = await resp.json();
+
+      dot.className = 'status-dot connected';
+      text.textContent = 'connected';
+
+
+      document.getElementById('totalRequests').textContent = fmt(data.total_requests || 0);
+      document.getElementById('totalPrompt').textContent = fmt(data.total_prompt_tokens || 0);
+      document.getElementById('totalCompletion').textContent = fmt(data.total_completion_tokens || 0);
+      document.getElementById('totalReasoning').textContent = fmt(data.total_reasoning_tokens || 0);
+      document.getElementById('totalTokens').textContent = fmt(data.total_tokens || 0);
+
+      // Record first non-zero reading as baseline
+      if (firstSeenTime === null && (data.total_tokens || 0) > 0) {
+        firstSeenTime = performance.now() / 1000;
+        firstSeenTokens = {
+          prompt: data.total_prompt_tokens || 0,
+          completion: data.total_completion_tokens || 0,
+          total: data.total_tokens || 0,
+        };
+      }
+
+      setRate('promptRate', data.total_prompt_tokens || 0, 'prompt');
+      setRate('completionRate', data.total_completion_tokens || 0, 'completion');
+      setRate('totalRate', data.total_tokens || 0, 'total');
+
+      renderModelTable(data.by_model || {});
+
+    } catch (e) {
+      dot.className = 'status-dot error';
+      text.textContent = e.message || 'error';
+    }
+  }
+
+  poll();
+  setInterval(poll, 1000);
+</script>
+</body>
+</html>

pyproject.toml

@@ -13,13 +13,14 @@ dependencies = [
     "filelock>=3.18.0",
     "rustworkx>=0.17.1",
     "huggingface-hub>=0.33.4",
+    "typer", # for huggingface-cli
     "psutil>=7.0.0",
     "loguru>=0.7.3",
     "exo_pyo3_bindings", # rust bindings
     "anyio==4.11.0",
     "mlx==0.30.3; sys_platform == 'darwin'",
     "mlx[cpu]==0.30.3; sys_platform == 'linux'",
-    "mlx-lm @ git+https://github.com/AlexCheema/mlx-lm.git@fix-transformers-5.0.0rc2",
+    "mlx-lm==0.30.5",
     "tiktoken>=0.12.0", # required for kimi k2 tokenizer
     "hypercorn>=0.18.0",
     "openai-harmony>=0.0.8",
@@ -34,6 +35,7 @@ dependencies = [
 exo-master = "exo.master.main:main"
 exo-worker = "exo.worker.main:main"
 exo = "exo.main:main"
+exo-eval = "bench.exo_eval:main"
 
 # dependencies only required for development
 [dependency-groups]
@@ -51,6 +53,9 @@ dev = [
 # cuda = [
 #     "mlx[cuda]==0.26.3",
 # ]
+eval = [
+    "lm_eval[api]",
+]
 
 ###
 # workspace configuration

src/exo/download/download_utils.py

@@ -121,11 +121,20 @@ async def ensure_models_dir() -> Path:
 
 
 async def delete_model(model_id: ModelId) -> bool:
-    model_dir = await ensure_models_dir() / model_id.normalize()
-    if not await aios.path.exists(model_dir):
-        return False
-    await asyncio.to_thread(shutil.rmtree, model_dir, ignore_errors=False)
-    return True
+    models_dir = await ensure_models_dir()
+    model_dir = models_dir / model_id.normalize()
+    cache_dir = models_dir / "caches" / model_id.normalize()
+
+    deleted = False
+    if await aios.path.exists(model_dir):
+        await asyncio.to_thread(shutil.rmtree, model_dir, ignore_errors=False)
+        deleted = True
+
+    # Also clear cache
+    if await aios.path.exists(cache_dir):
+        await asyncio.to_thread(shutil.rmtree, cache_dir, ignore_errors=False)
+
+    return deleted
 
 
 async def seed_models(seed_dir: str | Path):
@@ -151,16 +160,28 @@ async def fetch_file_list_with_cache(
     target_dir = (await ensure_models_dir()) / "caches" / model_id.normalize()
     await aios.makedirs(target_dir, exist_ok=True)
     cache_file = target_dir / f"{model_id.normalize()}--{revision}--file_list.json"
-    if await aios.path.exists(cache_file):
-        async with aiofiles.open(cache_file, "r") as f:
-            return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
-    file_list = await fetch_file_list_with_retry(
-        model_id, revision, recursive=recursive
-    )
-    await aios.makedirs(cache_file.parent, exist_ok=True)
-    async with aiofiles.open(cache_file, "w") as f:
-        await f.write(TypeAdapter(list[FileListEntry]).dump_json(file_list).decode())
-    return file_list
+
+    # Always try fresh first
+    try:
+        file_list = await fetch_file_list_with_retry(
+            model_id, revision, recursive=recursive
+        )
+        # Update cache with fresh data
+        async with aiofiles.open(cache_file, "w") as f:
+            await f.write(
+                TypeAdapter(list[FileListEntry]).dump_json(file_list).decode()
+            )
+        return file_list
+    except Exception as e:
+        # Fetch failed - try cache fallback
+        if await aios.path.exists(cache_file):
+            logger.warning(
+                f"Failed to fetch file list for {model_id}, using cached data: {e}"
+            )
+            async with aiofiles.open(cache_file, "r") as f:
+                return TypeAdapter(list[FileListEntry]).validate_json(await f.read())
+        # No cache available, propagate the error
+        raise
 
 
 async def fetch_file_list_with_retry(
@@ -332,8 +353,28 @@ async def _download_file(
     target_dir: Path,
     on_progress: Callable[[int, int, bool], None] = lambda _, __, ___: None,
 ) -> Path:
-    if await aios.path.exists(target_dir / path):
-        return target_dir / path
+    target_path = target_dir / path
+
+    if await aios.path.exists(target_path):
+        local_size = (await aios.stat(target_path)).st_size
+
+        # Try to verify against remote, but allow offline operation
+        try:
+            remote_size, _ = await file_meta(model_id, revision, path)
+            if local_size != remote_size:
+                logger.info(
+                    f"File {path} size mismatch (local={local_size}, remote={remote_size}), re-downloading"
+                )
+                await aios.remove(target_path)
+            else:
+                return target_path
+        except Exception as e:
+            # Offline or network error - trust local file
+            logger.debug(
+                f"Could not verify {path} against remote (offline?): {e}, using local file"
+            )
+            return target_path
+
     await aios.makedirs((target_dir / path).parent, exist_ok=True)
     length, etag = await file_meta(model_id, revision, path)
     remote_hash = etag[:-5] if etag.endswith("-gzip") else etag
@@ -542,17 +583,26 @@ async def download_shard(
     async def on_progress_wrapper(
         file: FileListEntry, curr_bytes: int, total_bytes: int, is_renamed: bool
     ) -> None:
-        start_time = (
-            file_progress[file.path].start_time
-            if file.path in file_progress
-            else time.time()
-        )
-        downloaded_this_session = (
-            file_progress[file.path].downloaded_this_session.in_bytes
-            + (curr_bytes - file_progress[file.path].downloaded.in_bytes)
-            if file.path in file_progress
-            else curr_bytes
+        previous_progress = file_progress.get(file.path)
+
+        # Detect re-download: curr_bytes < previous downloaded means file was deleted and restarted
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
         )
+
+        if is_redownload or previous_progress is None:
+            # Fresh download or re-download: reset tracking
+            start_time = time.time()
+            downloaded_this_session = curr_bytes
+        else:
+            # Continuing download: accumulate
+            start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
         speed = (
             downloaded_this_session / (time.time() - start_time)
             if time.time() - start_time > 0

src/exo/download/tests/test_download_verification.py

@@ -0,0 +1,451 @@
+"""Tests for download verification and cache behavior."""
+
+import time
+from collections.abc import AsyncIterator
+from datetime import timedelta
+from pathlib import Path
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import aiofiles
+import aiofiles.os as aios
+import pytest
+from pydantic import TypeAdapter
+
+from exo.download.download_utils import (
+    delete_model,
+    fetch_file_list_with_cache,
+)
+from exo.shared.types.common import ModelId
+from exo.shared.types.memory import Memory
+from exo.shared.types.worker.downloads import FileListEntry, RepoFileDownloadProgress
+
+
+@pytest.fixture
+def model_id() -> ModelId:
+    return ModelId("test-org/test-model")
+
+
+@pytest.fixture
+async def temp_models_dir(tmp_path: Path) -> AsyncIterator[Path]:
+    """Set up a temporary models directory for testing."""
+    models_dir = tmp_path / "models"
+    await aios.makedirs(models_dir, exist_ok=True)
+    with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+        yield models_dir
+
+
+class TestFileVerification:
+    """Tests for file size verification in _download_file."""
+
+    async def test_redownload_when_file_size_changes_upstream(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that files with mismatched sizes are re-downloaded."""
+        # Import inside test to allow patching
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file with wrong size
+        local_file = target_dir / "test.safetensors"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(b"local content")  # 13 bytes
+
+        remote_size = 1000  # Different from local
+        remote_hash = "abc123"
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                return_value=(remote_size, remote_hash),
+            ) as mock_file_meta,
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            # Set up mock HTTP response for re-download
+            mock_response = MagicMock()
+            mock_response.status = 200
+            mock_response.content.read = AsyncMock(  # pyright: ignore[reportAny]
+                side_effect=[b"x" * remote_size, b""]
+            )
+
+            mock_session = MagicMock()
+            mock_session.get.return_value.__aenter__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=mock_response
+            )
+            mock_session.get.return_value.__aexit__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=None
+            )
+            mock_session_factory.return_value.__aenter__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=mock_session
+            )
+            mock_session_factory.return_value.__aexit__ = AsyncMock(  # pyright: ignore[reportAny]
+                return_value=None
+            )
+
+            # Mock calc_hash to return the expected hash
+            with patch(
+                "exo.download.download_utils.calc_hash",
+                new_callable=AsyncMock,
+                return_value=remote_hash,
+            ):
+                await _download_file(model_id, "main", "test.safetensors", target_dir)
+
+            # file_meta should be called twice: once for verification, once for download
+            assert mock_file_meta.call_count == 2
+
+    async def test_skip_download_when_file_size_matches(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that files with matching sizes are not re-downloaded."""
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file
+        local_file = target_dir / "test.safetensors"
+        local_content = b"local content"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(local_content)
+
+        remote_size = len(local_content)  # Same as local
+        remote_hash = "abc123"
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                return_value=(remote_size, remote_hash),
+            ) as mock_file_meta,
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            result = await _download_file(
+                model_id, "main", "test.safetensors", target_dir
+            )
+
+            # Should return immediately without downloading
+            assert result == local_file
+            mock_file_meta.assert_called_once()
+            mock_session_factory.assert_not_called()
+
+    async def test_offline_fallback_uses_local_file(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that local files are used when network is unavailable."""
+        from exo.download.download_utils import (
+            _download_file,  # pyright: ignore[reportPrivateUsage]
+        )
+
+        target_dir = tmp_path / "downloads"
+        await aios.makedirs(target_dir, exist_ok=True)
+
+        # Create a local file
+        local_file = target_dir / "test.safetensors"
+        async with aiofiles.open(local_file, "wb") as f:
+            await f.write(b"local content")
+
+        with (
+            patch(
+                "exo.download.download_utils.file_meta",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+            patch(
+                "exo.download.download_utils.create_http_session"
+            ) as mock_session_factory,
+        ):
+            result = await _download_file(
+                model_id, "main", "test.safetensors", target_dir
+            )
+
+            # Should return local file without attempting download
+            assert result == local_file
+            mock_session_factory.assert_not_called()
+
+
+class TestFileListCache:
+    """Tests for file list caching behavior."""
+
+    async def test_fetch_fresh_and_update_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that fresh data is fetched and cache is updated."""
+        models_dir = tmp_path / "models"
+
+        file_list = [
+            FileListEntry(type="file", path="model.safetensors", size=1000),
+            FileListEntry(type="file", path="config.json", size=100),
+        ]
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                return_value=file_list,
+            ) as mock_fetch,
+        ):
+            result = await fetch_file_list_with_cache(model_id, "main")
+
+            assert result == file_list
+            mock_fetch.assert_called_once()
+
+            # Verify cache was written
+            cache_file = (
+                models_dir
+                / "caches"
+                / model_id.normalize()
+                / f"{model_id.normalize()}--main--file_list.json"
+            )
+            assert await aios.path.exists(cache_file)
+
+            async with aiofiles.open(cache_file, "r") as f:
+                cached_data = TypeAdapter(list[FileListEntry]).validate_json(
+                    await f.read()
+                )
+            assert cached_data == file_list
+
+    async def test_fallback_to_cache_when_fetch_fails(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that cached data is used when fetch fails."""
+        models_dir = tmp_path / "models"
+        cache_dir = models_dir / "caches" / model_id.normalize()
+        await aios.makedirs(cache_dir, exist_ok=True)
+
+        # Create cache file
+        cached_file_list = [
+            FileListEntry(type="file", path="model.safetensors", size=1000),
+        ]
+        cache_file = cache_dir / f"{model_id.normalize()}--main--file_list.json"
+        async with aiofiles.open(cache_file, "w") as f:
+            await f.write(
+                TypeAdapter(list[FileListEntry]).dump_json(cached_file_list).decode()
+            )
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+        ):
+            result = await fetch_file_list_with_cache(model_id, "main")
+
+            assert result == cached_file_list
+
+    async def test_error_propagates_when_no_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that errors propagate when fetch fails and no cache exists."""
+        models_dir = tmp_path / "models"
+
+        with (
+            patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir),
+            patch(
+                "exo.download.download_utils.fetch_file_list_with_retry",
+                new_callable=AsyncMock,
+                side_effect=Exception("Network error"),
+            ),
+            pytest.raises(Exception, match="Network error"),
+        ):
+            await fetch_file_list_with_cache(model_id, "main")
+
+
+class TestModelDeletion:
+    """Tests for model deletion including cache cleanup."""
+
+    async def test_delete_model_clears_cache(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test that deleting a model also deletes its cache."""
+        models_dir = tmp_path / "models"
+        model_dir = models_dir / model_id.normalize()
+        cache_dir = models_dir / "caches" / model_id.normalize()
+
+        # Create model and cache directories
+        await aios.makedirs(model_dir, exist_ok=True)
+        await aios.makedirs(cache_dir, exist_ok=True)
+
+        # Add some files
+        async with aiofiles.open(model_dir / "model.safetensors", "w") as f:
+            await f.write("model data")
+        async with aiofiles.open(cache_dir / "file_list.json", "w") as f:
+            await f.write("[]")
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            assert result is True
+            assert not await aios.path.exists(model_dir)
+            assert not await aios.path.exists(cache_dir)
+
+    async def test_delete_model_only_cache_exists(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test deleting when only cache exists (model already deleted)."""
+        models_dir = tmp_path / "models"
+        cache_dir = models_dir / "caches" / model_id.normalize()
+
+        # Only create cache directory
+        await aios.makedirs(cache_dir, exist_ok=True)
+        async with aiofiles.open(cache_dir / "file_list.json", "w") as f:
+            await f.write("[]")
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            # Returns False because model dir didn't exist
+            assert result is False
+            # But cache should still be cleaned up
+            assert not await aios.path.exists(cache_dir)
+
+    async def test_delete_nonexistent_model(
+        self, model_id: ModelId, tmp_path: Path
+    ) -> None:
+        """Test deleting a model that doesn't exist."""
+        models_dir = tmp_path / "models"
+        await aios.makedirs(models_dir, exist_ok=True)
+
+        with patch("exo.download.download_utils.EXO_MODELS_DIR", models_dir):
+            result = await delete_model(model_id)
+
+            assert result is False
+
+
+class TestProgressResetOnRedownload:
+    """Tests for progress tracking when files are re-downloaded."""
+
+    async def test_progress_resets_correctly_on_redownload(
+        self, model_id: ModelId
+    ) -> None:
+        """Test that progress tracking resets when a file is re-downloaded.
+
+        When a file is deleted and re-downloaded (due to size mismatch),
+        the progress tracking should reset rather than calculating negative
+        downloaded_this_session values.
+        """
+        # Simulate file_progress dict as it exists in download_shard
+        file_progress: dict[str, RepoFileDownloadProgress] = {}
+
+        # Initialize with old file progress (simulating existing large file)
+        old_file_size = 1_500_000_000  # 1.5 GB
+        file_progress["model.safetensors"] = RepoFileDownloadProgress(
+            repo_id=model_id,
+            repo_revision="main",
+            file_path="model.safetensors",
+            downloaded=Memory.from_bytes(old_file_size),
+            downloaded_this_session=Memory.from_bytes(0),
+            total=Memory.from_bytes(old_file_size),
+            speed=0,
+            eta=timedelta(0),
+            status="not_started",
+            start_time=time.time() - 10,  # Started 10 seconds ago
+        )
+
+        # Simulate the logic from on_progress_wrapper after re-download starts
+        # This is the exact logic from the fixed on_progress_wrapper
+        curr_bytes = 100_000  # 100 KB - new download just started
+        previous_progress = file_progress.get("model.safetensors")
+
+        # Detect re-download: curr_bytes < previous downloaded
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
+        )
+
+        if is_redownload or previous_progress is None:
+            # Fresh download or re-download: reset tracking
+            start_time = time.time()
+            downloaded_this_session = curr_bytes
+        else:
+            # Continuing download: accumulate
+            start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
+        # Key assertions
+        assert is_redownload is True, "Should detect re-download scenario"
+        assert downloaded_this_session == curr_bytes, (
+            "downloaded_this_session should equal curr_bytes on re-download"
+        )
+        assert downloaded_this_session > 0, (
+            "downloaded_this_session should be positive, not negative"
+        )
+
+        # Calculate speed (should be positive)
+        elapsed = time.time() - start_time
+        speed = downloaded_this_session / elapsed if elapsed > 0 else 0
+        assert speed >= 0, "Speed should be non-negative"
+
+    async def test_progress_accumulates_on_continuing_download(
+        self, model_id: ModelId
+    ) -> None:
+        """Test that progress accumulates correctly for continuing downloads.
+
+        When a download continues from where it left off (resume),
+        the progress should accumulate correctly.
+        """
+        file_progress: dict[str, RepoFileDownloadProgress] = {}
+
+        # Initialize with partial download progress
+        initial_downloaded = 500_000  # 500 KB already downloaded
+        start_time = time.time() - 5  # Started 5 seconds ago
+        file_progress["model.safetensors"] = RepoFileDownloadProgress(
+            repo_id=model_id,
+            repo_revision="main",
+            file_path="model.safetensors",
+            downloaded=Memory.from_bytes(initial_downloaded),
+            downloaded_this_session=Memory.from_bytes(initial_downloaded),
+            total=Memory.from_bytes(1_000_000),
+            speed=100_000,
+            eta=timedelta(seconds=5),
+            status="in_progress",
+            start_time=start_time,
+        )
+
+        # Progress callback with more bytes downloaded
+        curr_bytes = 600_000  # 600 KB - continuing download
+        previous_progress = file_progress.get("model.safetensors")
+
+        # This is NOT a re-download (curr_bytes > previous downloaded)
+        is_redownload = (
+            previous_progress is not None
+            and curr_bytes < previous_progress.downloaded.in_bytes
+        )
+
+        if is_redownload or previous_progress is None:
+            downloaded_this_session = curr_bytes
+            used_start_time = time.time()
+        else:
+            used_start_time = previous_progress.start_time
+            downloaded_this_session = (
+                previous_progress.downloaded_this_session.in_bytes
+                + (curr_bytes - previous_progress.downloaded.in_bytes)
+            )
+
+        # Key assertions
+        assert is_redownload is False, (
+            "Should NOT detect re-download for continuing download"
+        )
+        assert used_start_time == start_time, "Should preserve original start_time"
+        expected_session = initial_downloaded + (curr_bytes - initial_downloaded)
+        assert downloaded_this_session == expected_session, (
+            f"Should accumulate: {downloaded_this_session} == {expected_session}"
+        )
+        assert downloaded_this_session == 600_000, (
+            "downloaded_this_session should equal total downloaded so far"
+        )

src/exo/master/api.py

@@ -1,10 +1,11 @@
 import base64
 import contextlib
 import json
+import re
 import time
 from collections.abc import AsyncGenerator
 from http import HTTPStatus
-from typing import Annotated, Literal, cast
+from typing import Annotated, Any, Literal, cast
 from uuid import uuid4
 
 import anyio
@@ -42,6 +43,7 @@ from exo.shared.types.api import (
     ChatCompletionChoice,
     ChatCompletionMessage,
     ChatCompletionResponse,
+    CompletionTokensDetails,
     CreateInstanceParams,
     CreateInstanceResponse,
     DeleteDownloadResponse,
@@ -57,6 +59,8 @@ from exo.shared.types.api import (
     ImageGenerationTaskParams,
     ImageListItem,
     ImageListResponse,
+    Logprobs,
+    LogprobsContentItem,
     ModelList,
     ModelListModel,
     PlaceInstanceParams,
@@ -66,8 +70,10 @@ from exo.shared.types.api import (
     StartDownloadResponse,
     StreamingChoiceResponse,
     ToolCall,
+    Usage,
 )
 from exo.shared.types.chunks import (
+    CompletionChunk,
     ErrorChunk,
     ImageChunk,
     InputImageChunk,
@@ -107,14 +113,43 @@ from exo.utils.channels import Receiver, Sender, channel
 from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
+_THINK_TAG_RE = re.compile(r"<think>.*?</think>", re.DOTALL)
+
+
+def _strip_think_tags(text: str) -> str:
+    """Strip <think>...</think> blocks from response text.
+
+    These tags are an artifact of GPT-OSS channel parsing, not part of the
+    model's intended output. The OpenAI API content field should not contain them.
+    """
+    return _THINK_TAG_RE.sub("", text).lstrip()
+
 
 def _format_to_content_type(image_format: Literal["png", "jpeg", "webp"] | None) -> str:
     return f"image/{image_format or 'png'}"
 
 
+def _build_logprobs(chunk: TokenChunk) -> Logprobs:
+    """Convert flat logprob fields to OpenAI Logprobs format."""
+    return Logprobs(
+        content=[
+            LogprobsContentItem(
+                token=chunk.text,
+                logprob=chunk.logprob if chunk.logprob is not None else 0.0,
+                bytes=list(chunk.text.encode("utf-8")),
+                top_logprobs=chunk.top_logprobs or [],
+            )
+        ]
+    )
+
+
 def chunk_to_response(
     chunk: TokenChunk | ToolCallChunk, command_id: CommandId
 ) -> ChatCompletionResponse:
+    logprobs: Logprobs | None = None
+    if isinstance(chunk, TokenChunk) and chunk.logprob is not None:
+        logprobs = _build_logprobs(chunk)
+
     return ChatCompletionResponse(
         id=command_id,
         created=int(time.time()),
@@ -135,6 +170,7 @@ def chunk_to_response(
                         for i, tool in enumerate(chunk.tool_calls)
                     ],
                 ),
+                logprobs=logprobs,
                 finish_reason=chunk.finish_reason,
             )
         ],
@@ -197,7 +233,8 @@ class API:
         )
 
         self._chat_completion_queues: dict[
-            CommandId, Sender[TokenChunk | ErrorChunk | ToolCallChunk]
+            CommandId,
+            Sender[TokenChunk | ErrorChunk | ToolCallChunk | CompletionChunk],
         ] = {}
         self._image_generation_queues: dict[
             CommandId, Sender[ImageChunk | ErrorChunk]
@@ -205,6 +242,9 @@ class API:
         self._image_store = ImageStore(EXO_IMAGE_CACHE_DIR)
         self._tg: TaskGroup | None = None
 
+        # Accumulated usage stats per instance (keyed by model id)
+        self._usage_by_model: dict[str, dict[str, int]] = {}
+
     def reset(self, new_session_id: SessionId, result_clock: int):
         logger.info("Resetting API State")
         self.state = State()
@@ -271,6 +311,48 @@ class API:
         self.app.get("/events")(lambda: self._event_log)
         self.app.post("/download/start")(self.start_download)
         self.app.delete("/download/{node_id}/{model_id:path}")(self.delete_download)
+        self.app.get("/v1/usage")(self.get_usage)
+
+    def get_usage(self) -> dict[str, Any]:
+        """Return accumulated token usage per model instance."""
+        total_requests = 0
+        total_prompt = 0
+        total_completion = 0
+        total_reasoning = 0
+        for counters in self._usage_by_model.values():
+            total_requests += counters.get("requests", 0)
+            total_prompt += counters.get("prompt_tokens", 0)
+            total_completion += counters.get("completion_tokens", 0)
+            total_reasoning += counters.get("reasoning_tokens", 0)
+        return {
+            "total_requests": total_requests,
+            "total_prompt_tokens": total_prompt,
+            "total_completion_tokens": total_completion,
+            "total_reasoning_tokens": total_reasoning,
+            "total_tokens": total_prompt + total_completion,
+            "by_model": self._usage_by_model,
+        }
+
+    def _accumulate_usage(
+        self,
+        model: str,
+        prompt_tokens: int,
+        completion_tokens: int,
+        reasoning_tokens: int,
+    ) -> None:
+        """Accumulate usage stats for a model instance."""
+        if model not in self._usage_by_model:
+            self._usage_by_model[model] = {
+                "requests": 0,
+                "prompt_tokens": 0,
+                "completion_tokens": 0,
+                "reasoning_tokens": 0,
+            }
+        counters = self._usage_by_model[model]
+        counters["requests"] += 1
+        counters["prompt_tokens"] += prompt_tokens
+        counters["completion_tokens"] += completion_tokens
+        counters["reasoning_tokens"] += reasoning_tokens
 
     async def place_instance(self, payload: PlaceInstanceParams):
         command = PlaceInstance(
@@ -492,29 +574,37 @@ class API:
         )
 
     async def _chat_chunk_stream(
-        self, command_id: CommandId
-    ) -> AsyncGenerator[ErrorChunk | ToolCallChunk | TokenChunk, None]:
-        """Yield `TokenChunk`s for a given command until completion."""
+        self, command_id: CommandId, timeout: float = 60000.0
+    ) -> AsyncGenerator[TokenChunk | ErrorChunk | ToolCallChunk, None]:
+        """Yield `TokenChunk`s for a given command until completion.
+
+        Args:
+            timeout: Max seconds to wait for the next chunk before aborting.
+        """
 
         try:
             self._chat_completion_queues[command_id], recv = channel[
-                ErrorChunk | ToolCallChunk | TokenChunk
+                TokenChunk | ErrorChunk | ToolCallChunk
             ]()
 
             with recv as token_chunks:
-                async for chunk in token_chunks:
-                    yield chunk
-                    if chunk.finish_reason is not None:
-                        break
+                with anyio.fail_after(timeout):
+                    async for chunk in token_chunks:
+                        yield chunk
+                        if chunk.finish_reason is not None:
+                            break
 
         except anyio.get_cancelled_exc_class():
-            # TODO: TaskCancelled
-            """
-            self.command_sender.send_nowait(
-                ForwarderCommand(origin=self.node_id, command=command)
-            )
-            """
             raise
+        except TimeoutError:
+            logger.warning(
+                f"Chat completion timed out after {timeout}s (command_id={command_id})"
+            )
+            yield ErrorChunk(
+                model=ModelId("unknown"),
+                finish_reason="error",
+                error_message=f"Request timed out after {timeout}s",
+            )
         finally:
             command = TaskFinished(finished_command_id=command_id)
             await self._send(command)
@@ -528,7 +618,7 @@ class API:
 
         async for chunk in self._chat_chunk_stream(command_id):
             assert not isinstance(chunk, ImageChunk)
-            if chunk.finish_reason == "error":
+            if isinstance(chunk, ErrorChunk):
                 error_response = ErrorResponse(
                     error=ErrorInfo(
                         message=chunk.error_message or "Internal server error",
@@ -548,6 +638,15 @@ class API:
             yield f"data: {chunk_response.model_dump_json()}\n\n"
 
             if chunk.finish_reason is not None:
+                # Accumulate usage stats from the final chunk
+                if isinstance(chunk, TokenChunk) and chunk.stats is not None:
+                    s = chunk.stats
+                    self._accumulate_usage(
+                        model=chunk.model,
+                        prompt_tokens=s.prompt_tokens,
+                        completion_tokens=s.generation_tokens,
+                        reasoning_tokens=s.reasoning_tokens,
+                    )
                 yield "data: [DONE]\n\n"
 
     async def _collect_chat_completion(
@@ -557,10 +656,14 @@ class API:
 
         text_parts: list[str] = []
         tool_calls: list[ToolCall] = []
+        logprobs_items: list[LogprobsContentItem] = []
         model: str | None = None
         finish_reason: FinishReason | None = None
+        stats: GenerationStats | None = None
 
         async for chunk in self._chat_chunk_stream(command_id):
+            # Skip CompletionChunk - it's for the legacy completions API
+
             if isinstance(chunk, ErrorChunk):
                 raise HTTPException(
                     status_code=500,
@@ -572,6 +675,16 @@ class API:
 
             if isinstance(chunk, TokenChunk):
                 text_parts.append(chunk.text)
+                if chunk.stats is not None:
+                    stats = chunk.stats
+                if chunk.logprob is not None:
+                    lp = _build_logprobs(chunk)
+                    if lp.content:
+                        if len(lp.content) != 1:
+                            logger.warning(
+                                f"Expected 1 logprobs content item per chunk, got {len(lp.content)}"
+                            )
+                        logprobs_items.append(lp.content[0])
 
             if isinstance(chunk, ToolCallChunk):
                 tool_calls.extend(
@@ -586,9 +699,33 @@ class API:
             if chunk.finish_reason is not None:
                 finish_reason = chunk.finish_reason
 
-        combined_text = "".join(text_parts)
+        combined_text = _strip_think_tags("".join(text_parts))
         assert model is not None
 
+        logprobs: Logprobs | None = None
+        if logprobs_items:
+            logprobs = Logprobs(content=logprobs_items)
+
+        usage: Usage | None = None
+        if stats is not None:
+            completion_tokens = stats.generation_tokens
+            usage = Usage(
+                prompt_tokens=stats.prompt_tokens,
+                completion_tokens=completion_tokens,
+                total_tokens=stats.prompt_tokens + completion_tokens,
+                completion_tokens_details=CompletionTokensDetails(
+                    reasoning_tokens=stats.reasoning_tokens,
+                )
+                if stats.reasoning_tokens > 0
+                else None,
+            )
+            self._accumulate_usage(
+                model=model or "unknown",
+                prompt_tokens=stats.prompt_tokens,
+                completion_tokens=completion_tokens,
+                reasoning_tokens=stats.reasoning_tokens,
+            )
+
         return ChatCompletionResponse(
             id=command_id,
             created=int(time.time()),
@@ -601,9 +738,11 @@ class API:
                         content=combined_text,
                         tool_calls=tool_calls,
                     ),
+                    logprobs=logprobs,
                     finish_reason=finish_reason,
                 )
             ],
+            usage=usage,
         )
 
     async def _collect_chat_completion_with_stats(
@@ -617,7 +756,7 @@ class API:
         stats: GenerationStats | None = None
 
         async for chunk in self._chat_chunk_stream(command_id):
-            if chunk.finish_reason == "error":
+            if isinstance(chunk, ErrorChunk):
                 raise HTTPException(
                     status_code=500,
                     detail=chunk.error_message or "Internal server error",
@@ -628,6 +767,7 @@ class API:
 
             if isinstance(chunk, TokenChunk):
                 text_parts.append(chunk.text)
+                stats = chunk.stats or stats
 
             if isinstance(chunk, ToolCallChunk):
                 tool_calls.extend(
@@ -638,13 +778,12 @@ class API:
                     )
                     for i, tool in enumerate(chunk.tool_calls)
                 )
-
-            stats = chunk.stats or stats
+                stats = chunk.stats or stats
 
             if chunk.finish_reason is not None:
                 finish_reason = chunk.finish_reason
 
-        combined_text = "".join(text_parts)
+        combined_text = _strip_think_tags("".join(text_parts))
         assert model is not None
 
         resp = BenchChatCompletionResponse(
@@ -695,7 +834,14 @@ class API:
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id)
+        try:
+            return await self._collect_chat_completion(command.command_id)
+        except BaseException:
+            # Ensure task cleanup if handler is cancelled before _chat_chunk_stream's finally runs
+            with contextlib.suppress(Exception):
+                await self._send(TaskFinished(finished_command_id=command.command_id))
+            self._chat_completion_queues.pop(command.command_id, None)
+            raise
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionTaskParams

src/exo/master/main.py

@@ -13,6 +13,7 @@ from exo.master.placement import (
 from exo.shared.apply import apply
 from exo.shared.types.commands import (
     ChatCompletion,
+    Completion,
     CreateInstance,
     DeleteInstance,
     ForwarderCommand,
@@ -40,6 +41,9 @@ from exo.shared.types.state import State
 from exo.shared.types.tasks import (
     ChatCompletion as ChatCompletionTask,
 )
+from exo.shared.types.tasks import (
+    Completion as CompletionTask,
+)
 from exo.shared.types.tasks import (
     ImageEdits as ImageEditsTask,
 )
@@ -158,6 +162,48 @@ class Master:
                                 )
                             )
 
+                            self.command_task_mapping[command.command_id] = task_id
+                        case Completion():
+                            for instance in self.state.instances.values():
+                                if (
+                                    instance.shard_assignments.model_id
+                                    == command.request_params.model
+                                ):
+                                    task_count = sum(
+                                        1
+                                        for task in self.state.tasks.values()
+                                        if task.instance_id == instance.instance_id
+                                    )
+                                    instance_task_counts[instance.instance_id] = (
+                                        task_count
+                                    )
+
+                            if not instance_task_counts:
+                                raise ValueError(
+                                    f"No instance found for model {command.request_params.model}"
+                                )
+
+                            available_instance_ids = sorted(
+                                instance_task_counts.keys(),
+                                key=lambda instance_id: instance_task_counts[
+                                    instance_id
+                                ],
+                            )
+
+                            task_id = TaskId()
+                            generated_events.append(
+                                TaskCreated(
+                                    task_id=task_id,
+                                    task=CompletionTask(
+                                        task_id=task_id,
+                                        command_id=command.command_id,
+                                        instance_id=available_instance_ids[0],
+                                        task_status=TaskStatus.Pending,
+                                        task_params=command.request_params,
+                                    ),
+                                )
+                            )
+
                             self.command_task_mapping[command.command_id] = task_id
                         case ImageGeneration():
                             for instance in self.state.instances.values():
@@ -279,17 +325,15 @@ class Master:
                                 )
                             )
                         case TaskFinished():
-                            generated_events.append(
-                                TaskDeleted(
-                                    task_id=self.command_task_mapping[
-                                        command.finished_command_id
-                                    ]
-                                )
+                            task_id = self.command_task_mapping.pop(
+                                command.finished_command_id, None
                             )
-                            if command.finished_command_id in self.command_task_mapping:
-                                del self.command_task_mapping[
-                                    command.finished_command_id
-                                ]
+                            if task_id is not None:
+                                generated_events.append(TaskDeleted(task_id=task_id))
+                            else:
+                                logger.debug(
+                                    f"TaskFinished for unknown command_id={command.finished_command_id} (already cleaned up)"
+                                )
                         case RequestEventLog():
                             # We should just be able to send everything, since other buffers will ignore old messages
                             for i in range(command.since_idx, len(self._event_log)):

src/exo/shared/models/model_cards.py

@@ -413,9 +413,9 @@ MODEL_CARDS: dict[str, ModelCard] = {
     ),
 }
 
-_IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
+_IMAGE_BASE_MODEL_CARDS: dict[str, ModelCard] = {
     "flux1-schnell": ModelCard(
-        model_id=ModelId("black-forest-labs/FLUX.1-schnell"),
+        model_id=ModelId("exolabs/FLUX.1-schnell"),
         storage_size=Memory.from_bytes(23782357120 + 9524621312),
         n_layers=57,
         hidden_size=1,
@@ -428,7 +428,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
                 storage_size=Memory.from_kb(0),
                 n_layers=12,
                 can_shard=False,
-                safetensors_index_filename=None,  # Single file
+                safetensors_index_filename=None,
             ),
             ComponentInfo(
                 component_name="text_encoder_2",
@@ -442,7 +442,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
                 component_name="transformer",
                 component_path="transformer/",
                 storage_size=Memory.from_bytes(23782357120),
-                n_layers=57,  # 19 transformer_blocks + 38 single_transformer_blocks
+                n_layers=57,
                 can_shard=True,
                 safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
             ),
@@ -457,7 +457,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
         ],
     ),
     "flux1-dev": ModelCard(
-        model_id=ModelId("black-forest-labs/FLUX.1-dev"),
+        model_id=ModelId("exolabs/FLUX.1-dev"),
         storage_size=Memory.from_bytes(23782357120 + 9524621312),
         n_layers=57,
         hidden_size=1,
@@ -470,7 +470,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
                 storage_size=Memory.from_kb(0),
                 n_layers=12,
                 can_shard=False,
-                safetensors_index_filename=None,  # Single file
+                safetensors_index_filename=None,
             ),
             ComponentInfo(
                 component_name="text_encoder_2",
@@ -484,7 +484,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
                 component_name="transformer",
                 component_path="transformer/",
                 storage_size=Memory.from_bytes(23802816640),
-                n_layers=57,  # 19 transformer_blocks + 38 single_transformer_blocks
+                n_layers=57,
                 can_shard=True,
                 safetensors_index_filename="diffusion_pytorch_model.safetensors.index.json",
             ),
@@ -499,7 +499,7 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
         ],
     ),
     "flux1-krea-dev": ModelCard(
-        model_id=ModelId("black-forest-labs/FLUX.1-Krea-dev"),
+        model_id=ModelId("exolabs/FLUX.1-Krea-dev"),
         storage_size=Memory.from_bytes(23802816640 + 9524621312),  # Same as dev
         n_layers=57,
         hidden_size=1,
@@ -541,9 +541,9 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
         ],
     ),
     "qwen-image": ModelCard(
-        model_id=ModelId("Qwen/Qwen-Image"),
+        model_id=ModelId("exolabs/Qwen-Image"),
         storage_size=Memory.from_bytes(16584333312 + 40860802176),
-        n_layers=60,  # Qwen has 60 transformer blocks (all joint-style)
+        n_layers=60,
         hidden_size=1,
         supports_tensor=False,
         tasks=[ModelTask.TextToImage],
@@ -551,10 +551,10 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
             ComponentInfo(
                 component_name="text_encoder",
                 component_path="text_encoder/",
-                storage_size=Memory.from_kb(16584333312),
+                storage_size=Memory.from_bytes(16584333312),
                 n_layers=12,
                 can_shard=False,
-                safetensors_index_filename=None,  # Single file
+                safetensors_index_filename=None,
             ),
             ComponentInfo(
                 component_name="transformer",
@@ -575,9 +575,9 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
         ],
     ),
     "qwen-image-edit-2509": ModelCard(
-        model_id=ModelId("Qwen/Qwen-Image-Edit-2509"),
+        model_id=ModelId("exolabs/Qwen-Image-Edit-2509"),
         storage_size=Memory.from_bytes(16584333312 + 40860802176),
-        n_layers=60,  # Qwen has 60 transformer blocks (all joint-style)
+        n_layers=60,
         hidden_size=1,
         supports_tensor=False,
         tasks=[ModelTask.ImageToImage],
@@ -585,10 +585,10 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
             ComponentInfo(
                 component_name="text_encoder",
                 component_path="text_encoder/",
-                storage_size=Memory.from_kb(16584333312),
+                storage_size=Memory.from_bytes(16584333312),
                 n_layers=12,
                 can_shard=False,
-                safetensors_index_filename=None,  # Single file
+                safetensors_index_filename=None,
             ),
             ComponentInfo(
                 component_name="transformer",
@@ -610,6 +610,92 @@ _IMAGE_MODEL_CARDS: dict[str, ModelCard] = {
     ),
 }
 
+
+def _generate_image_model_quant_variants(
+    base_name: str,
+    base_card: ModelCard,
+) -> dict[str, ModelCard]:
+    """Create quantized variants of an image model card.
+
+    Only the transformer component is quantized; text encoders stay at bf16.
+    Sizes are calculated exactly from the base card's component sizes.
+    """
+    if base_card.components is None:
+        raise ValueError(f"Image model {base_name} must have components defined")
+
+    # quantizations = [8, 6, 5, 4, 3]
+    quantizations = [8, 4]
+
+    num_transformer_bytes = next(
+        c.storage_size.in_bytes
+        for c in base_card.components
+        if c.component_name == "transformer"
+    )
+
+    transformer_bytes = Memory.from_bytes(num_transformer_bytes)
+
+    remaining_bytes = Memory.from_bytes(
+        sum(
+            c.storage_size.in_bytes
+            for c in base_card.components
+            if c.component_name != "transformer"
+        )
+    )
+
+    def with_transformer_size(new_size: Memory) -> list[ComponentInfo]:
+        assert base_card.components is not None
+        return [
+            ComponentInfo(
+                component_name=c.component_name,
+                component_path=c.component_path,
+                storage_size=new_size
+                if c.component_name == "transformer"
+                else c.storage_size,
+                n_layers=c.n_layers,
+                can_shard=c.can_shard,
+                safetensors_index_filename=c.safetensors_index_filename,
+            )
+            for c in base_card.components
+        ]
+
+    variants = {
+        base_name: ModelCard(
+            model_id=base_card.model_id,
+            storage_size=transformer_bytes + remaining_bytes,
+            n_layers=base_card.n_layers,
+            hidden_size=base_card.hidden_size,
+            supports_tensor=base_card.supports_tensor,
+            tasks=base_card.tasks,
+            components=with_transformer_size(transformer_bytes),
+        )
+    }
+
+    for quant in quantizations:
+        quant_transformer_bytes = Memory.from_bytes(
+            (num_transformer_bytes * quant) // 16
+        )
+        total_bytes = remaining_bytes + quant_transformer_bytes
+
+        model_id = ModelId(base_card.model_id + f"-{quant}bit")
+
+        variants[f"{base_name}-{quant}bit"] = ModelCard(
+            model_id=model_id,
+            storage_size=total_bytes,
+            n_layers=base_card.n_layers,
+            hidden_size=base_card.hidden_size,
+            supports_tensor=base_card.supports_tensor,
+            tasks=base_card.tasks,
+            components=with_transformer_size(quant_transformer_bytes),
+        )
+
+    return variants
+
+
+_image_model_cards: dict[str, ModelCard] = {}
+for _base_name, _base_card in _IMAGE_BASE_MODEL_CARDS.items():
+    _image_model_cards |= _generate_image_model_quant_variants(_base_name, _base_card)
+_IMAGE_MODEL_CARDS = _image_model_cards
+
 if EXO_ENABLE_IMAGE_MODELS:
     MODEL_CARDS.update(_IMAGE_MODEL_CARDS)
 

src/exo/shared/types/api.py

@@ -98,6 +98,8 @@ class LogprobsContentItem(BaseModel):
 
 class Logprobs(BaseModel):
     content: list[LogprobsContentItem] | None = None
+    # This will always be null for open source models, but exists for OpenAI API
+    refusal: list[LogprobsContentItem] | None = None
 
 
 class PromptTokensDetails(BaseModel):
@@ -150,6 +152,7 @@ class GenerationStats(BaseModel):
     generation_tps: float
     prompt_tokens: int
     generation_tokens: int
+    reasoning_tokens: int = 0
     peak_memory_usage: Memory
 
 
@@ -170,6 +173,52 @@ class BenchChatCompletionResponse(ChatCompletionResponse):
     generation_stats: GenerationStats | None = None
 
 
+# Legacy Completions API types (for lm_eval compatibility)
+class CompletionLogprobs(BaseModel):
+    """Logprobs in the legacy completions format."""
+
+    tokens: list[str]
+    token_logprobs: list[float | None]
+    top_logprobs: list[dict[str, float]]
+    text_offset: list[int]
+
+
+class CompletionChoice(BaseModel):
+    text: str
+    index: int
+    logprobs: CompletionLogprobs | None = None
+    finish_reason: FinishReason | None = None
+
+
+class CompletionResponse(BaseModel):
+    id: str
+    object: Literal["text_completion"] = "text_completion"
+    created: int
+    model: str
+    choices: list[CompletionChoice]
+    usage: Usage | None = None
+
+
+class CompletionTaskParams(BaseModel):
+    """Parameters for the legacy /v1/completions endpoint."""
+
+    model: str
+    # Prompt can be: string, list of strings, list of token IDs, or list of token ID lists
+    prompt: str | list[str] | list[int] | list[list[int]]
+    max_tokens: int | None = 16
+    temperature: float | None = 1.0
+    top_p: float | None = 1.0
+    n: int | None = 1
+    stream: bool = False
+    logprobs: int | None = None
+    echo: bool = False
+    stop: str | list[str] | None = None
+    presence_penalty: float | None = None
+    frequency_penalty: float | None = None
+    seed: int | None = None
+    user: str | None = None
+
+
 class ChatCompletionTaskParams(BaseModel):
     model: str
     frequency_penalty: float | None = None

src/exo/shared/types/chunks.py

@@ -2,7 +2,7 @@ from collections.abc import Generator
 from typing import Any, Literal
 
 from exo.shared.models.model_cards import ModelId
-from exo.shared.types.api import GenerationStats, ImageGenerationStats
+from exo.shared.types.api import GenerationStats, ImageGenerationStats, TopLogprobItem
 from exo.utils.pydantic_ext import TaggedModel
 
 from .api import FinishReason
@@ -17,6 +17,8 @@ class BaseChunk(TaggedModel):
 class TokenChunk(BaseChunk):
     text: str
     token_id: int
+    logprob: float | None = None
+    top_logprobs: list[TopLogprobItem] | None = None
     finish_reason: Literal["stop", "length", "content_filter"] | None = None
     stats: GenerationStats | None = None
 
@@ -32,6 +34,17 @@ class ToolCallChunk(BaseChunk):
     stats: GenerationStats | None = None
 
 
+class CompletionChunk(BaseChunk):
+    """Chunk for legacy completions API with full logprobs for all tokens."""
+
+    text: str
+    tokens: list[str]
+    token_logprobs: list[float | None]
+    top_logprobs: list[dict[str, float]]
+    text_offset: list[int]
+    finish_reason: FinishReason | None = None
+
+
 class ImageChunk(BaseChunk):
     data: str
     chunk_index: int
@@ -67,4 +80,4 @@ class InputImageChunk(BaseChunk):
                 yield name, value
 
 
-GenerationChunk = TokenChunk | ImageChunk | ToolCallChunk | ErrorChunk
+GenerationChunk = TokenChunk | CompletionChunk | ImageChunk | ToolCallChunk | ErrorChunk

src/exo/shared/types/commands.py

@@ -3,6 +3,7 @@ from pydantic import Field
 from exo.shared.models.model_cards import ModelCard, ModelId
 from exo.shared.types.api import (
     ChatCompletionTaskParams,
+    CompletionTaskParams,
     ImageEditsInternalParams,
     ImageGenerationTaskParams,
 )
@@ -25,6 +26,12 @@ class ChatCompletion(BaseCommand):
     request_params: ChatCompletionTaskParams
 
 
+class Completion(BaseCommand):
+    """Legacy completions API command for scoring/generation."""
+
+    request_params: CompletionTaskParams
+
+
 class ImageGeneration(BaseCommand):
     request_params: ImageGenerationTaskParams
 
@@ -79,6 +86,7 @@ Command = (
     TestCommand
     | RequestEventLog
     | ChatCompletion
+    | Completion
     | ImageGeneration
     | ImageEdits
     | PlaceInstance

src/exo/shared/types/mlx.py

@@ -0,0 +1,12 @@
+"""Shared types for MLX-related functionality."""
+
+from collections.abc import Sequence
+
+from mlx_lm.models.cache import (
+    KVCache,
+    QuantizedKVCache,
+    RotatingKVCache,
+)
+
+# This list contains one cache entry per transformer layer
+KVCacheType = Sequence[KVCache | RotatingKVCache | QuantizedKVCache]

src/exo/shared/types/tasks.py

@@ -4,6 +4,7 @@ from pydantic import Field
 
 from exo.shared.types.api import (
     ChatCompletionTaskParams,
+    CompletionTaskParams,
     ImageEditsInternalParams,
     ImageGenerationTaskParams,
 )
@@ -60,6 +61,16 @@ class ChatCompletion(BaseTask):  # emitted by Master
     error_message: str | None = Field(default=None)
 
 
+class Completion(BaseTask):
+    """Legacy completions task for scoring tokens with echo=True."""
+
+    command_id: CommandId
+    task_params: CompletionTaskParams
+
+    error_type: str | None = Field(default=None)
+    error_message: str | None = Field(default=None)
+
+
 class ImageGeneration(BaseTask):  # emitted by Master
     command_id: CommandId
     task_params: ImageGenerationTaskParams
@@ -87,6 +98,7 @@ Task = (
     | LoadModel
     | StartWarmup
     | ChatCompletion
+    | Completion
     | ImageGeneration
     | ImageEdits
     | Shutdown

src/exo/shared/types/worker/runner_response.py

@@ -6,6 +6,7 @@ from exo.shared.types.api import (
     GenerationStats,
     ImageGenerationStats,
     ToolCallItem,
+    TopLogprobItem,
 )
 from exo.utils.pydantic_ext import TaggedModel
 
@@ -14,14 +15,11 @@ class BaseRunnerResponse(TaggedModel):
     pass
 
 
-class TokenizedResponse(BaseRunnerResponse):
-    prompt_tokens: int
-
-
 class GenerationResponse(BaseRunnerResponse):
     text: str
     token: int
-    # logprobs: list[float] | None = None # too big. we can change to be top-k
+    logprob: float | None = None
+    top_logprobs: list[TopLogprobItem] | None = None
     finish_reason: FinishReason | None = None
     stats: GenerationStats | None = None
 

src/exo/utils/channels.py

@@ -194,6 +194,22 @@ class MpReceiver[T]:
                 raise EndOfStream from None
             return item
 
+    def receive_with_timeout(self, timeout: float) -> T | None:
+        """Receive with timeout, returns None if no message within timeout."""
+        if self._state.closed.is_set():
+            raise ClosedResourceError
+
+        try:
+            item = self._state.buffer.get(block=True, timeout=timeout)
+            if isinstance(item, _MpEndOfStream):
+                self.close()
+                raise EndOfStream
+            return item
+        except Empty:
+            return None
+        except ValueError as e:
+            raise ClosedResourceError from e
+
     # nb: this function will not cancel particularly well
     async def receive_async(self) -> T:
         return await to_thread.run_sync(self.receive, limiter=CapacityLimiter(1))

src/exo/worker/engines/mlx/auto_parallel.py

@@ -13,12 +13,17 @@ from mlx.nn.layers.distributed import (
     shard_linear,
     sum_gradients,
 )
+from mlx_lm.models.base import (
+    scaled_dot_product_attention,  # pyright: ignore[reportUnknownVariableType]
+)
 from mlx_lm.models.deepseek_v3 import DeepseekV3MLP
 from mlx_lm.models.deepseek_v3 import Model as DeepseekV3Model
 from mlx_lm.models.deepseek_v32 import DeepseekV32MLP
 from mlx_lm.models.deepseek_v32 import Model as DeepseekV32Model
 from mlx_lm.models.glm4_moe import Model as Glm4MoeModel
 from mlx_lm.models.glm4_moe import MoE
+from mlx_lm.models.glm4_moe_lite import Glm4MoeLiteDecoderLayer, Glm4MoeLiteMLP
+from mlx_lm.models.glm4_moe_lite import Model as GLM4MoeLiteModel
 from mlx_lm.models.gpt_oss import GptOssMoeModel
 from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.models.llama import Model as LlamaModel
@@ -27,7 +32,8 @@ from mlx_lm.models.ministral3 import Model as Ministral3Model
 from mlx_lm.models.qwen3_moe import Model as Qwen3MoeModel
 from mlx_lm.models.qwen3_moe import Qwen3MoeSparseMoeBlock
 from mlx_lm.models.qwen3_next import Model as Qwen3NextModel
-from mlx_lm.models.qwen3_next import Qwen3NextSparseMoeBlock
+from mlx_lm.models.qwen3_next import Qwen3NextDecoderLayer, Qwen3NextSparseMoeBlock
+from transformers.models.qwen3.modeling_qwen3 import Qwen3DecoderLayer
 
 from exo.shared.logging import logger
 from exo.shared.types.worker.shards import PipelineShardMetadata
@@ -100,6 +106,16 @@ class CustomMlxLayer(nn.Module):
                 return getattr(original_layer, name)
 
 
+class EvalCheckpointLayer(CustomMlxLayer):
+    """Wraps a layer to force evaluation of its output, breaking up the computation graph
+    to prevent Metal command buffer timeouts with large batches in pipeline parallel."""
+
+    def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
+        output = self.original_layer(x, *args, **kwargs)
+        mx.eval(output)
+        return output
+
+
 class PipelineFirstLayer(CustomMlxLayer):
     def __init__(
         self,
@@ -137,14 +153,20 @@ class PipelineLastLayer(CustomMlxLayer):
         ).arguments.get("cache", None)
 
         output: mx.array = self.original_layer(x, *args, **kwargs)
+        mx.eval(output)
 
         if self.r != self.s - 1:
             output = mx.distributed.send(
                 output, (self.r + 1) % self.s, group=self.group
             )
+            mx.async_eval(output)
             if cache is not None:
                 cache.keys = mx.depends(cache.keys, output)  # type: ignore[reportUnknownMemberType]
 
+        output = mx.distributed.all_gather(output, group=self.group)[
+            -output.shape[0] :
+        ]  # type :ignore
+
         return output
 
 
@@ -195,6 +217,9 @@ def pipeline_auto_parallel(
 
     layers = layers[start_layer:end_layer]
     layers[0] = PipelineFirstLayer(layers[0], device_rank, group=group)
+    # Wrap intermediate layers with eval checkpoints to prevent GPU timeout
+    for i in range(1, len(layers) - 1):
+        layers[i] = EvalCheckpointLayer(layers[i])
     layers[-1] = PipelineLastLayer(
         layers[-1],
         device_rank,
@@ -248,14 +273,14 @@ def patch_pipeline_model[T](model: T, group: mx.distributed.Group) -> T:
             "cache", None
         )
 
+        # Evaluate logits before all_gather to break the computation graph
+        # and prevent Metal command buffer timeouts with large batches
+        mx.eval(logits)
+
         # Add dependency to last cache entry to ensure distributed ops are evaluated
         if cache is not None:
             cache[-1].state = mx.depends(cache[-1].state, logits)  # type: ignore
 
-        logits = mx.distributed.all_gather(logits, group=group)[
-            -logits.shape[0] :
-        ]  # type :ignore
-
         return logits
 
     cls.__call__ = patched_call
@@ -334,15 +359,7 @@ def tensor_auto_parallel(
         group=group,
     )
 
-    if hasattr(model, "shard") and not isinstance(model, GptOssModel):
-        try:
-            model.shard(group)  # type: ignore
-            return patch_tensor_model(model)
-        except (AttributeError, TypeError, NameError):
-            pass
-
     if isinstance(model, (LlamaModel, Ministral3Model)):
-        logger.warning("shouldn't be hit - upstream sharding exists")
         tensor_parallel_sharding_strategy = LlamaShardingStrategy(
             group,
             all_to_sharded_linear,
@@ -351,7 +368,6 @@ def tensor_auto_parallel(
             sharded_to_all_linear_in_place,
         )
     elif isinstance(model, (DeepseekV3Model, DeepseekV32Model)):
-        logger.warning("shouldn't be hit - upstream sharding exists")
         tensor_parallel_sharding_strategy = DeepSeekShardingStrategy(
             group,
             all_to_sharded_linear,
@@ -367,6 +383,14 @@ def tensor_auto_parallel(
             all_to_sharded_linear_in_place,
             sharded_to_all_linear_in_place,
         )
+    elif isinstance(model, GLM4MoeLiteModel):
+        tensor_parallel_sharding_strategy = GLM4MoeLiteShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
     elif isinstance(model, (Qwen3MoeModel, Glm4MoeModel, Qwen3NextModel)):
         tensor_parallel_sharding_strategy = QwenShardingStrategy(
             group,
@@ -441,7 +465,7 @@ class LlamaShardingStrategy(TensorParallelShardingStrategy):
             layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
             layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
             layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
-
+            mx.eval(layer)
         return model
 
 
@@ -516,6 +540,8 @@ class DeepSeekShardingStrategy(TensorParallelShardingStrategy):
                 layer.mlp = ShardedDeepseekV3MoE(layer.mlp)  # type: ignore
                 layer.mlp.sharding_group = self.group
 
+            mx.eval(layer)
+
         return model
 
 
@@ -533,6 +559,158 @@ class ShardedDeepseekV3MoE(CustomMlxLayer):
         return y
 
 
+class GLM4MoeLiteShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(GLM4MoeLiteModel, model)
+        for layer in model.layers:  # type: ignore
+            layer = cast(Glm4MoeLiteDecoderLayer, layer)
+            eval_with_timeout(
+                layer.parameters(),
+                timeout_seconds / len(model.layers),  # type: ignore
+                on_timeout,
+            )
+            if layer.self_attn.q_lora_rank is None:  # type: ignore
+                layer.self_attn.q_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_proj
+                )
+            else:
+                layer.self_attn.q_b_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_b_proj
+                )
+
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+            layer.self_attn.num_heads //= self.N
+
+            # Logic from upstream mlx
+            num_heads = layer.self_attn.num_heads
+            sh = self.group.rank() * num_heads
+            eh = sh + num_heads
+
+            def shard_heads(w: mx.array, sh: int = sh, eh: int = eh) -> mx.array:
+                return w[sh:eh]
+
+            layer.self_attn.embed_q.apply(shard_heads)
+            layer.self_attn.unembed_out.apply(shard_heads)
+
+            if isinstance(layer.mlp, Glm4MoeLiteMLP):
+                layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
+                layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
+                layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
+
+            else:
+                if getattr(layer.mlp, "shared_experts", None) is not None:
+                    self.all_to_sharded_linear_in_place(
+                        layer.mlp.shared_experts.gate_proj
+                    )
+                    self.sharded_to_all_linear_in_place(
+                        layer.mlp.shared_experts.down_proj
+                    )
+                    self.all_to_sharded_linear_in_place(
+                        layer.mlp.shared_experts.up_proj
+                    )
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.gate_proj)
+                self.sharded_to_all_linear_in_place(layer.mlp.switch_mlp.down_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.up_proj)
+                layer.mlp = ShardedGLM4MoeLiteMoE(layer.mlp)  # type: ignore
+                layer.mlp.sharding_group = self.group  # type: ignore
+            mx.eval(layer)
+
+        return model
+
+
+class ShardedGLM4MoeLiteMoE(CustomMlxLayer):
+    def __init__(self, layer: _LayerCallable):
+        super().__init__(layer)
+        self.sharding_group: mx.distributed.Group | None = None
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+        y = self.original_layer.__call__(x)
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+        return y
+
+
+class WrappedMiniMaxAttention(CustomMlxLayer):
+    def __init__(self, layer: _LayerCallable, group: mx.distributed.Group):
+        super().__init__(layer)
+        self.group = group
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: mx.array | Any = None,
+        cache: Any | None = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+
+        if getattr(self, "use_qk_norm", False):
+            q_dim = queries.shape[-1]  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            k_dim = keys.shape[-1]  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            N = self.group.size()
+
+            qk = mx.concatenate([queries, keys], axis=-1)  # (B, L, q_dim + k_dim)
+            qk = mx.distributed.all_gather(
+                qk, group=self.group
+            )  # (N*B, L, q_dim + k_dim)
+
+            # Reshape to separate rank contributions: (N, B, L, q_dim + k_dim)
+            # Then transpose to (B, L, N, q_dim + k_dim) and merge N into feature dim
+            qk = qk.reshape(N, B, L, q_dim + k_dim).transpose(1, 2, 0, 3)  # pyright: ignore[reportUnknownMemberType,reportUnknownArgumentType]
+            queries = qk[..., :q_dim].reshape(
+                B, L, -1
+            )  # (B, L, N * q_dim)  # pyright: ignore[reportUnknownMemberType]
+            keys = qk[..., q_dim:].reshape(
+                B, L, -1
+            )  # (B, L, N * k_dim)  # pyright: ignore[reportUnknownMemberType]
+
+            queries = self.q_norm(queries)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            keys = self.k_norm(keys)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+
+            # Split back and take this rank's portion
+            queries = mx.split(queries, N, axis=-1)[self.group.rank()]
+            keys = mx.split(keys, N, axis=-1)[self.group.rank()]
+
+        queries = queries.reshape(B, L, self.num_attention_heads, -1).transpose(  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType,reportUnknownArgumentType]
+            0, 2, 1, 3
+        )
+        keys = keys.reshape(B, L, self.num_key_value_heads, -1).transpose(  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType,reportUnknownArgumentType]
+            0, 2, 1, 3
+        )
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            0, 2, 1, 3
+        )
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType,reportAny]
+            keys = self.rope(keys, offset=cache.offset)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType,reportAny]
+            keys, values = cache.update_and_fetch(keys, values)  # pyright: ignore[reportAny]
+        else:
+            queries = self.rope(queries)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            keys = self.rope(keys)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+
+        output = scaled_dot_product_attention(
+            queries,
+            keys,
+            values,
+            cache=cache,
+            scale=self.scale,
+            mask=mask,  # pyright: ignore[reportUnknownMemberType,reportUnknownArgumentType]
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)  # pyright: ignore[reportUnknownMemberType]
+
+        return self.o_proj(output)  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+
+
 class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
     def shard_model(
         self,
@@ -550,9 +728,12 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
             layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
             layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
             layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+
             layer.self_attn.num_attention_heads //= self.N
             layer.self_attn.num_key_value_heads //= self.N
 
+            layer.self_attn = WrappedMiniMaxAttention(layer.self_attn, self.group)  # pyright: ignore[reportAttributeAccessIssue,reportArgumentType]
+
             # Shard the MoE. Shard in place since the MoE should be responsible
             # for aggregating the results.
             self.all_to_sharded_linear_in_place(
@@ -566,7 +747,7 @@ class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
             )
             layer.block_sparse_moe = ShardedQwenMoE(layer.block_sparse_moe)  # pyright: ignore[reportAttributeAccessIssue, reportArgumentType]
             layer.block_sparse_moe.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
-
+            mx.eval(layer)
         return model
 
 
@@ -577,18 +758,32 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
         timeout_seconds: float,
         on_timeout: TimeoutCallback | None,
     ) -> nn.Module:
-        model = cast(Qwen3MoeModel, model)
+        model = cast(Qwen3MoeModel | Qwen3NextModel, model)
         for layer in model.layers:
             eval_with_timeout(
                 layer.parameters(), timeout_seconds / len(model.layers), on_timeout
             )
             # Shard the self attention
-            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
-            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
-            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
-            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
-            layer.self_attn.n_heads //= self.N
-            layer.self_attn.n_kv_heads //= self.N
+            if isinstance(layer, Qwen3DecoderLayer) or hasattr(layer, "self_attn"):
+                layer.self_attn.q_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_proj
+                )
+                layer.self_attn.k_proj = self.all_to_sharded_linear(
+                    layer.self_attn.k_proj
+                )
+                layer.self_attn.v_proj = self.all_to_sharded_linear(
+                    layer.self_attn.v_proj
+                )
+                layer.self_attn.o_proj = self.sharded_to_all_linear(
+                    layer.self_attn.o_proj
+                )
+                layer.self_attn.n_heads //= self.N
+                layer.self_attn.n_kv_heads //= self.N
+            else:
+                assert isinstance(layer, Qwen3NextDecoderLayer) and hasattr(
+                    layer, "linear_attn"
+                )
+                # These layers are fast so we don't shard. This may change in future.
 
             # Shard the MoE. Shard in place since the MoE should be responsible
             # for aggregating the results.
@@ -607,6 +802,7 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
                 layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
                 layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
 
+            mx.eval(layer)
         return model
 
 
@@ -661,7 +857,7 @@ class GptOssShardingStrategy(TensorParallelShardingStrategy):
 
             layer.mlp = ShardedGptOssMoE(layer.mlp)  # type: ignore
             layer.mlp.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
-
+            mx.eval(layer)
         return model
 
 

src/exo/worker/engines/mlx/cache.py

@@ -1,39 +1,81 @@
-# type: ignore
-# TODO: Fix this file, including types!
+import os
 from copy import deepcopy
-from typing import Callable
+from typing import Any, cast
 
 import mlx.core as mx
-from mlx_lm import stream_generate
-from mlx_lm.models.cache import _BaseCache, trim_prompt_cache
+from mlx_lm.models.cache import (
+    KVCache,
+    QuantizedKVCache,
+    RotatingKVCache,
+    trim_prompt_cache,
+)
+from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
+from exo.shared.types.mlx import KVCacheType
 from exo.worker.engines.mlx import Model
-from exo.worker.engines.mlx.constants import KEEP_KV_SIZE, KV_BITS, KV_GROUP_SIZE
-from exo.worker.engines.mlx.utils_mlx import make_kv_cache
+from exo.worker.engines.mlx.constants import CACHE_GROUP_SIZE, KV_CACHE_BITS
+from exo.worker.runner.bootstrap import logger
+
+# Fraction of device memory above which LRU eviction kicks in
+_DEFAULT_MEMORY_THRESHOLD = 0.85
+_MEMORY_THRESHOLD = float(
+    os.environ.get("EXO_MEMORY_THRESHOLD", _DEFAULT_MEMORY_THRESHOLD)
+)
 
 
 class KVPrefixCache:
-    def __init__(self):
-        # Only one prefix cache per runner.
+    def __init__(self, tokenizer: TokenizerWrapper):
         self.prompts: list[mx.array] = []  # mx array of tokens (ints)
-        self.caches: list[list[_BaseCache]] = []
+        self.caches: list[KVCacheType] = []
+        self._last_used: list[int] = []  # monotonic counter of last access per entry
+        self._access_counter: int = 0
+        self._tokenizer: TokenizerWrapper = tokenizer
 
-    def add_kv_cache(
-        self, tokenizer: TokenizerWrapper, prompt: str, cache: list[_BaseCache]
-    ):
-        tokenized_prompt = self.encode_prompt(tokenizer, prompt)
+    def clear(self):
+        """Clear all cached prompts and caches."""
+        self.prompts.clear()
+        self.caches.clear()
+        self._last_used.clear()
+
+    def add_kv_cache(self, prompt: str, cache: KVCacheType):
+        """Add a new cache entry. Evicts LRU entries if memory is high."""
+        self._evict_if_needed()
+        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
         self.prompts.append(tokenized_prompt)
         self.caches.append(deepcopy(cache))
+        self._access_counter += 1
+        self._last_used.append(self._access_counter)
+        logger.info(f"KV cache added: {len(tokenized_prompt)} tokens")
+
+    def update_kv_cache(
+        self,
+        index: int,
+        prompt: str,
+        cache: KVCacheType,
+    ):
+        """Update an existing cache entry in-place."""
+        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
+        self.prompts[index] = tokenized_prompt
+        self.caches[index] = deepcopy(cache)
+        self._access_counter += 1
+        self._last_used[index] = self._access_counter
+        logger.info(f"KV cache updated (index {index}): {len(tokenized_prompt)} tokens")
 
     def get_kv_cache(
         self,
         model: Model,
-        tokenizer: TokenizerWrapper,
-        sampler: Callable[[mx.array], mx.array],
         prompt: str,
-    ) -> list[_BaseCache]:
-        tokenized_prompt = self.encode_prompt(tokenizer, prompt)
+    ) -> tuple[KVCacheType, mx.array, int | None]:
+        """Get KV cache for prompt, returning remaining tokens to prefill.
+
+        Returns:
+            Tuple of (cache, remaining_tokens, matched_index) where:
+            - cache: KV cache to use for generation
+            - remaining_tokens: tokens that still need prefilling
+            - matched_index: index of the matched entry (None if no match)
+        """
+        tokenized_prompt = encode_prompt(self._tokenizer, prompt)
         max_length = len(tokenized_prompt)
 
         best_snapshot_index, best_snapshot_length = None, 0
@@ -42,63 +84,127 @@ class KVPrefixCache:
             length = _get_prefix_length(tokenized_prompt, cached_prompt)
 
             if length == max_length:
-                return self.caches[i]
+                # Exact match - cached prompt starts with our entire prompt
+                # Trim cache to prompt length - 1, return last token for stream_generate
+                prompt_cache = deepcopy(self.caches[i])
+                cached_length = _cache_length(self.caches[i])
+                tokens_to_trim = cached_length - (max_length - 1)
+                if tokens_to_trim > 0:
+                    trim_prompt_cache(cast(list[Any], prompt_cache), tokens_to_trim)
+                self._access_counter += 1
+                self._last_used[i] = self._access_counter
+                logger.info(f"KV cache exact match: {max_length} tokens (instant)")
+                return prompt_cache, tokenized_prompt[-1:], i
 
             if length > best_snapshot_length:
                 best_snapshot_index, best_snapshot_length = i, length
 
         if best_snapshot_index is not None:
-            prompt_cache = deepcopy(self.caches[best_snapshot_index])
-            trim_prompt_cache(prompt_cache, max_length - best_snapshot_length)
-            tokenized_prompt = tokenized_prompt[best_snapshot_index:]
-
-        else:
-            prompt_cache = make_kv_cache(
-                model,
-                # max_kv_size=MAX_KV_SIZE,
-                # keep=KEEP_KV_SIZE
+            new_tokens = max_length - best_snapshot_length
+            logger.info(
+                f"KV cache prefix match: {best_snapshot_length}/{max_length} tokens "
+                f"(reusing {best_snapshot_length}, need to prefill {new_tokens})"
             )
 
-        prefill(model, tokenizer, sampler, tokenized_prompt, prompt_cache)
+            prompt_cache = deepcopy(self.caches[best_snapshot_index])
 
-        return prompt_cache
+            # Trim removes tokens from the end, so we trim (cached_length - prefix_length) to keep the prefix
+            cached_length = _cache_length(self.caches[best_snapshot_index])
+            tokens_to_trim = cached_length - best_snapshot_length
+            if tokens_to_trim > 0:
+                trim_prompt_cache(cast(list[Any], prompt_cache), tokens_to_trim)
 
-    def encode_prompt(self, tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
-        add_special_tokens = tokenizer.bos_token is None or not prompt.startswith(
-            tokenizer.bos_token
-        )
-        tokenized_prompt = tokenizer.encode(
-            prompt, add_special_tokens=add_special_tokens
-        )
-        return mx.array(tokenized_prompt)
+            self._access_counter += 1
+            self._last_used[best_snapshot_index] = self._access_counter
+            remaining_tokens = tokenized_prompt[best_snapshot_length:]
+            return prompt_cache, remaining_tokens, best_snapshot_index
+
+        else:
+            prompt_cache = make_kv_cache(model)
+            if len(self.prompts) == 0:
+                logger.info(f"KV cache empty, need to prefill {max_length} tokens")
+            else:
+                logger.info(
+                    f"KV cache no prefix match, need to prefill {max_length} tokens"
+                )
+
+            return prompt_cache, tokenized_prompt, None
+
+    def _evict_if_needed(self):
+        """Evict least recently used entries while memory pressure is high."""
+        if len(self.caches) == 0:
+            return
+
+        active: int = mx.metal.get_active_memory()
+        limit = int(mx.metal.device_info()["max_recommended_working_set_size"])
+        if active < limit * _MEMORY_THRESHOLD:
+            return
+
+        # Evict LRU entries until below threshold or only one entry left
+        while len(self.caches) > 0:
+            lru_index = self._last_used.index(min(self._last_used))
+            evicted_tokens = len(self.prompts[lru_index])
+            self.prompts.pop(lru_index)
+            self.caches.pop(lru_index)
+            self._last_used.pop(lru_index)
+            logger.info(
+                f"KV cache evicted LRU entry ({evicted_tokens} tokens) due to memory pressure"
+            )
+
+            active = mx.metal.get_active_memory()
+            if active < limit * _MEMORY_THRESHOLD:
+                break
+
+
+def encode_prompt(tokenizer: TokenizerWrapper, prompt: str) -> mx.array:
+    """Encode a prompt string to token array.
+
+    For chat-templated prompts (which have their own structure markers like
+    <|im_user|>, <|im_middle|>, etc.), we should NOT add BOS/EOS tokens as
+    that would corrupt the prompt structure.
+    """
+    # Chat templates define their own structure - don't add BOS/EOS
+    tokenized_prompt = tokenizer.encode(prompt, add_special_tokens=False)
+    return mx.array(tokenized_prompt)
+
+
+def _cache_length(cache: KVCacheType) -> int:
+    """Get the number of tokens in a KV cache."""
+    # Use .offset attribute which all cache types have (len() not implemented in older QuantizedKVCache)
+    return max(c.offset for c in cache)  # type: ignore
 
 
 def _get_prefix_length(prompt: mx.array, cached_prompt: mx.array) -> int:
-    n = min(int(prompt.shape[0]), int(cached_prompt.shape[0]), KEEP_KV_SIZE)
+    """Find the length of the common prefix between two token arrays."""
+    n = min(int(prompt.shape[0]), int(cached_prompt.shape[0]))
     if n == 0:
         return 0
 
-    equal = (prompt[:n] == cached_prompt[:n]).astype(mx.int32)
+    equal = mx.equal(prompt[:n], cached_prompt[:n]).astype(mx.int32)
     prefix_mask = mx.cumprod(equal)  # stays 1 until first mismatch, then 0 forever
     return int(mx.sum(prefix_mask).item())
 
 
-def prefill(
-    model: Model,
-    tokenizer: TokenizerWrapper,
-    sampler: Callable[[mx.array], mx.array],
-    prompt: mx.array,
-    cache: list[_BaseCache],
-) -> None:
-    for _ in stream_generate(
-        model=model,
-        tokenizer=tokenizer,
-        prompt=prompt,
-        max_tokens=0,
-        sampler=sampler,
-        prompt_cache=cache,
-        prefill_step_size=2048,
-        kv_group_size=KV_GROUP_SIZE,
-        kv_bits=KV_BITS,
-    ):
-        pass
+def make_kv_cache(
+    model: Model, max_kv_size: int | None = None, keep: int = 0
+) -> KVCacheType:
+    assert hasattr(model, "layers")
+
+    # TODO: Do this for all models
+    if hasattr(model, "make_cache") and isinstance(model, GptOssModel):
+        logger.info("Using MLX LM's make cache")
+        return model.make_cache()  # type: ignore
+
+    if max_kv_size is None:
+        if KV_CACHE_BITS is None:
+            logger.info("Using default KV cache")
+            return [KVCache() for _ in model.layers]
+        else:
+            logger.info("Using quantized KV cache")
+            return [
+                QuantizedKVCache(group_size=CACHE_GROUP_SIZE, bits=KV_CACHE_BITS)
+                for _ in model.layers
+            ]
+    else:
+        logger.info(f"Using rotating KV cache with {max_kv_size=} with {keep=}")
+        return [RotatingKVCache(max_size=max_kv_size, keep=keep) for _ in model.layers]

src/exo/worker/engines/mlx/constants.py

@@ -4,7 +4,7 @@
 KV_GROUP_SIZE: int | None = 32
 KV_BITS: int | None = None
 ATTENTION_KV_BITS: int | None = 4
-MAX_TOKENS: int = 8192
+MAX_TOKENS: int = 32168
 MAX_KV_SIZE: int | None = 3200
 KEEP_KV_SIZE: int | None = 1600
 QUANTIZE_MODEL_MODE: str | None = "affine"

src/exo/worker/engines/mlx/generator/generate.py

@@ -1,48 +1,92 @@
+import time
 from typing import Any, Callable, Generator, cast, get_args
 
 import mlx.core as mx
 from mlx_lm.generate import stream_generate
-from mlx_lm.models.cache import KVCache
+from mlx_lm.models.cache import KVCache, trim_prompt_cache
 from mlx_lm.sample_utils import make_sampler
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
-# from exo.engines.mlx.cache import KVPrefixCache
 from exo.shared.types.api import (
     BenchChatCompletionTaskParams,
     ChatCompletionMessage,
     FinishReason,
     GenerationStats,
+    TopLogprobItem,
 )
 from exo.shared.types.memory import Memory
+from exo.shared.types.mlx import KVCacheType
 from exo.shared.types.tasks import ChatCompletionTaskParams
 from exo.shared.types.worker.runner_response import (
     GenerationResponse,
 )
 from exo.worker.engines.mlx import Model
+from exo.worker.engines.mlx.cache import KVPrefixCache, encode_prompt, make_kv_cache
 from exo.worker.engines.mlx.constants import KV_BITS, KV_GROUP_SIZE, MAX_TOKENS
 from exo.worker.engines.mlx.utils_mlx import (
     apply_chat_template,
-    make_kv_cache,
     mx_barrier,
 )
 from exo.worker.runner.bootstrap import logger
 
 generation_stream = mx.new_stream(mx.default_device())
 
+_MIN_PREFIX_HIT_TO_UPDATE = 1000
 
-def maybe_quantize_kv_cache(
-    prompt_cache: list[KVCache | Any],
-    quantized_kv_start: int,
-    kv_group_size: int,
-    kv_bits: int | None,
-) -> None:
-    if kv_bits is None:
-        return
-    for e, c in enumerate(prompt_cache):
-        if (
-            hasattr(c, "to_quantized") and c.offset >= quantized_kv_start  # type: ignore
-        ):
-            prompt_cache[e] = c.to_quantized(group_size=kv_group_size, bits=kv_bits)
+
+def prefill(
+    model: Model,
+    tokenizer: TokenizerWrapper,
+    sampler: Callable[[mx.array], mx.array],
+    prompt_tokens: mx.array,
+    cache: KVCacheType,
+) -> float:
+    """Prefill the KV cache with prompt tokens.
+
+    This runs the model over the prompt tokens to populate the cache,
+    then trims off the extra generated token.
+
+    Returns:
+        tokens_per_sec
+    """
+    num_tokens = len(prompt_tokens)
+    if num_tokens == 0:
+        return 0.0
+
+    logger.debug(f"Prefilling {num_tokens} tokens...")
+    start_time = time.perf_counter()
+
+    def progress_callback(processed: int, total: int) -> None:
+        elapsed = time.time() - start_time
+        tok_per_sec = processed / elapsed if elapsed > 0 else 0
+        logger.debug(
+            f"Prefill progress: {processed}/{total} tokens ({tok_per_sec:.1f} tok/s)"
+        )
+
+    # Use max_tokens=1 because max_tokens=0 does not work.
+    # We just throw away the generated token - we only care about filling the cache
+    for _ in stream_generate(
+        model=model,
+        tokenizer=tokenizer,
+        prompt=prompt_tokens,
+        max_tokens=1,
+        sampler=sampler,
+        prompt_cache=cache,
+        prefill_step_size=2048,
+        kv_group_size=KV_GROUP_SIZE,
+        kv_bits=KV_BITS,
+        prompt_progress_callback=progress_callback,
+    ):
+        break  # Stop after first iteration - cache is now filled
+    trim_prompt_cache(cast(list[Any], cache), 1)
+
+    elapsed = time.perf_counter() - start_time
+    tokens_per_sec = num_tokens / elapsed if elapsed > 0 else 0.0
+    logger.debug(
+        f"Prefill complete: {num_tokens} tokens in {elapsed:.2f}s "
+        f"({tokens_per_sec:.1f} tok/s)"
+    )
+    return tokens_per_sec
 
 
 def warmup_inference(
@@ -115,11 +159,212 @@ def eos_ids_from_tokenizer(tokenizer: TokenizerWrapper) -> list[int]:
     return eos
 
 
+def extract_top_logprobs(
+    logprobs_array: mx.array,
+    selected_token: int,
+    tokenizer: TokenizerWrapper,
+    top_k: int | None,
+) -> tuple[float, list[TopLogprobItem]]:
+    """Extract the selected token's logprob and top-k alternatives.
+
+    top k an be set to None to return all the logprobs
+    """
+    selected_logprob = float(logprobs_array[selected_token].item())
+
+    if top_k == 0:
+        return selected_logprob, []
+
+    vocab_size = logprobs_array.shape[0]
+
+    if top_k is None:
+        sorted_indices = mx.argsort(-logprobs_array)
+        mx.eval(sorted_indices)
+        indices_list: list[int] = cast(list[int], sorted_indices.tolist())
+    else:
+        k = min(top_k, vocab_size)
+        top_indices = mx.argpartition(-logprobs_array, kth=k - 1)[:k]
+        top_logprobs_values = logprobs_array[top_indices]
+        sorted_order = mx.argsort(-top_logprobs_values)
+        top_indices = top_indices[sorted_order]
+        mx.eval(top_indices)
+        indices_list = cast(list[int], top_indices.tolist())
+
+    top_logprob_items: list[TopLogprobItem] = []
+    for token_id in indices_list:
+        logprob_value = float(logprobs_array[token_id].item())
+        token_str = tokenizer.decode([token_id])
+
+        top_logprob_items.append(
+            TopLogprobItem(
+                token=token_str,
+                logprob=logprob_value,
+                bytes=list(token_str.encode("utf-8")),
+            )
+        )
+
+    return selected_logprob, top_logprob_items
+
+
+def score_tokens(
+    model: Model,
+    tokenizer: TokenizerWrapper,
+    tokens: list[int],
+    top_k: int | None = None,
+) -> list[tuple[float, list[TopLogprobItem]]]:
+    """Score a sequence of tokens, returning logprobs for each token.
+
+    This is used for the completions API with echo=True, where we need
+    logprobs for the prompt tokens (not just generated tokens).
+
+    Args:
+        model: The MLX model.
+        tokenizer: The tokenizer.
+        tokens: List of token IDs to score.
+        top_k: Number of top logprobs to return per position.
+               If None, returns all logprobs.
+
+    Returns:
+        List of (token_logprob, top_logprobs) tuples for each token position.
+        The first position has no logprob (no previous context), so returns (0.0, []).
+    """
+    if len(tokens) == 0:
+        return []
+
+    # First token has no previous context to condition on
+    results: list[tuple[float, list[TopLogprobItem]]] = [(0.0, [])]
+
+    if len(tokens) == 1:
+        return results
+
+    # Create an empty KV cache for the forward pass
+    cache = make_kv_cache(model=model)
+
+    # Convert to MLX array and run forward pass
+    input_tokens = mx.array(tokens[:-1])[None]  # All tokens except last, batched
+
+    # Run the model to get logits for all positions
+    # The model returns logits with shape [1, seq_len, vocab_size]
+    logits: mx.array = model(input_tokens, cache=cast(list[KVCache], cache))
+    logits = logits.squeeze(0)  # Shape: [seq_len, vocab_size]
+
+    # Convert to log probabilities
+    logprobs_all: mx.array = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
+
+    mx.eval(logprobs_all)
+
+    # For each position, extract the logprob of the actual next token
+    for i in range(len(tokens) - 1):
+        next_token = tokens[i + 1]
+        logprobs_at_position: mx.array = logprobs_all[i]
+
+        logprob, top_logprobs_items = extract_top_logprobs(
+            logprobs_array=logprobs_at_position,
+            selected_token=next_token,
+            tokenizer=tokenizer,
+            top_k=top_k,
+        )
+        results.append((logprob, top_logprobs_items))
+
+    return results
+
+
+def score_tokens_batched(
+    model: Model,
+    tokenizer: TokenizerWrapper,
+    token_sequences: list[list[int]],
+    top_k: int | None = None,
+) -> list[list[tuple[float, list[TopLogprobItem]]]]:
+    """Score multiple token sequences in a single batched forward pass.
+
+    This is significantly faster than calling score_tokens() multiple times
+    because it batches the forward pass across all sequences.
+
+    Args:
+        model: The MLX model.
+        tokenizer: The tokenizer.
+        token_sequences: List of token ID sequences to score.
+        top_k: Number of top logprobs to return per position.
+
+    Returns:
+        List of results for each sequence. Each result is a list of
+        (token_logprob, top_logprobs) tuples for each token position.
+    """
+    if not token_sequences:
+        return []
+
+    # Handle empty sequences and single-token sequences
+    results: list[list[tuple[float, list[TopLogprobItem]]]] = []
+    non_empty_indices: list[int] = []
+    non_empty_sequences: list[list[int]] = []
+
+    for i, tokens in enumerate(token_sequences):
+        if len(tokens) == 0:
+            results.append([])
+        elif len(tokens) == 1:
+            results.append([(0.0, [])])
+        else:
+            results.append([])  # Placeholder, will be filled later
+            non_empty_indices.append(i)
+            non_empty_sequences.append(tokens)
+
+    if not non_empty_sequences:
+        return results
+
+    # Find max sequence length (excluding last token since we predict it)
+    max_len = max(len(seq) - 1 for seq in non_empty_sequences)
+
+    # Get pad token (use eos_token_id or 0)
+    pad_token_id = getattr(tokenizer, "pad_token_id", None)
+    if pad_token_id is None:
+        pad_token_id = getattr(tokenizer, "eos_token_id", 0)
+
+    # Pad sequences and create attention mask
+    batch_size = len(non_empty_sequences)
+    padded_inputs = mx.full((batch_size, max_len), pad_token_id, dtype=mx.int32)
+    seq_lengths: list[int] = []
+
+    for i, tokens in enumerate(non_empty_sequences):
+        input_len = len(tokens) - 1  # Exclude last token
+        padded_inputs[i, :input_len] = mx.array(tokens[:-1], dtype=mx.int32)
+        seq_lengths.append(input_len)
+
+    # Run batched forward pass (no KV cache for scoring)
+    # The model accepts [batch_size, seq_len] and returns [batch_size, seq_len, vocab_size]
+    logits = model(padded_inputs, cache=None)
+
+    # Convert to log probabilities - logits shape: [batch, seq_len, vocab]
+    logprobs_all = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
+    mx.eval(logprobs_all)
+
+    # Extract results for each sequence
+    for batch_idx, (orig_idx, tokens, seq_len) in enumerate(
+        zip(non_empty_indices, non_empty_sequences, seq_lengths, strict=True)
+    ):
+        seq_results: list[tuple[float, list[TopLogprobItem]]] = [(0.0, [])]
+
+        for pos in range(seq_len):
+            next_token = tokens[pos + 1]
+            logprobs_at_position: mx.array = logprobs_all[batch_idx, pos]
+
+            logprob, top_logprobs_items = extract_top_logprobs(
+                logprobs_array=logprobs_at_position,
+                selected_token=next_token,
+                tokenizer=tokenizer,
+                top_k=top_k,
+            )
+            seq_results.append((logprob, top_logprobs_items))
+
+        results[orig_idx] = seq_results
+
+    return results
+
+
 def mlx_generate(
     model: Model,
     tokenizer: TokenizerWrapper,
     task: ChatCompletionTaskParams,
     prompt: str,
+    kv_prefix_cache: KVPrefixCache | None = None,
 ) -> Generator[GenerationResponse]:
     # Ensure that generation stats only contains peak memory for this generation
     mx.reset_peak_memory()
@@ -131,7 +376,22 @@ def mlx_generate(
     if task.seed is not None:
         mx.random.seed(task.seed)
 
-    caches = make_kv_cache(model=model)
+    # Do not use the prefix cache if we are trying to do benchmarks.
+    if is_bench:
+        kv_prefix_cache = None
+
+    # Use prefix cache if available, otherwise create fresh cache
+    prefix_hit_length = 0
+    matched_index: int | None = None
+    if kv_prefix_cache is None:
+        caches = make_kv_cache(model=model)
+        prompt_tokens = encode_prompt(tokenizer, prompt)
+    else:
+        caches, prompt_tokens, matched_index = kv_prefix_cache.get_kv_cache(
+            model, prompt
+        )
+        all_prompt_tokens = encode_prompt(tokenizer, prompt)
+        prefix_hit_length = len(all_prompt_tokens) - len(prompt_tokens)
 
     logits_processors: list[Callable[[mx.array, mx.array], mx.array]] = []
     if is_bench:
@@ -144,11 +404,23 @@ def mlx_generate(
         top_p=task.top_p if task.top_p is not None else 1.0,
     )
 
+    # Prefill cache with all tokens except the last one
+    prefill_tps = prefill(model, tokenizer, sampler, prompt_tokens[:-1], caches)
+
+    # stream_generate starts from the last token
+    last_token = prompt_tokens[-1:]
+
+    # Determine if we need logprobs
+    should_extract_logprobs = task.logprobs is True
+    top_k = task.top_logprobs if task.top_logprobs is not None else 0
+
     max_tokens = task.max_tokens or MAX_TOKENS
+    generated_text_parts: list[str] = []
+    generation_start_time = time.perf_counter()
     for out in stream_generate(
         model=model,
         tokenizer=tokenizer,
-        prompt=prompt,
+        prompt=last_token,
         max_tokens=max_tokens,
         sampler=sampler,
         logits_processors=logits_processors,
@@ -158,12 +430,13 @@ def mlx_generate(
         kv_group_size=KV_GROUP_SIZE,
         kv_bits=KV_BITS,
     ):
+        generated_text_parts.append(out.text)
         logger.info(out.text)
 
         stats: GenerationStats | None = None
         if out.finish_reason is not None:
             stats = GenerationStats(
-                prompt_tps=float(out.prompt_tps),
+                prompt_tps=float(prefill_tps or out.prompt_tps),
                 generation_tps=float(out.generation_tps),
                 prompt_tokens=int(out.prompt_tokens),
                 generation_tokens=int(out.generation_tokens),
@@ -177,14 +450,47 @@ def mlx_generate(
                     f"Model generated unexpected finish_reason: {out.finish_reason}"
                 )
 
+        # Extract logprobs if requested
+        logprob: float | None = None
+        top_logprobs: list[TopLogprobItem] | None = None
+        if should_extract_logprobs:
+            logprob, top_logprobs = extract_top_logprobs(
+                logprobs_array=out.logprobs,
+                selected_token=out.token,
+                tokenizer=tokenizer,
+                top_k=top_k,
+            )
+
         yield GenerationResponse(
             text=out.text,
             token=out.token,
+            logprob=logprob,
+            top_logprobs=top_logprobs,
             finish_reason=cast(FinishReason | None, out.finish_reason),
             stats=stats,
         )
 
         if out.finish_reason is not None:
+            # Log generation stats
+            generation_elapsed = time.perf_counter() - generation_start_time
+            generated_tokens = len(generated_text_parts)
+            generation_tps = (
+                generated_tokens / generation_elapsed if generation_elapsed > 0 else 0.0
+            )
+            logger.debug(
+                f"Generation complete: prefill {prompt_tokens} tokens @ "
+                f"{prefill_tps:.1f} tok/s, generated {generated_tokens} tokens @ "
+                f"{generation_tps:.1f} tok/s"
+            )
+            if kv_prefix_cache is not None:
+                full_prompt = prompt + "".join(generated_text_parts)
+                if (
+                    matched_index is not None
+                    and prefix_hit_length >= _MIN_PREFIX_HIT_TO_UPDATE
+                ):
+                    kv_prefix_cache.update_kv_cache(matched_index, full_prompt, caches)
+                else:
+                    kv_prefix_cache.add_kv_cache(full_prompt, caches)
             break
 
         # TODO: Do we want an mx_barrier?

src/exo/worker/engines/mlx/utils_mlx.py

@@ -18,15 +18,12 @@ try:
 except ImportError:
     pass  # transformers < 5.0 or bytes_to_unicode not available
 
-from mlx_lm.models.cache import KVCache, QuantizedKVCache, RotatingKVCache
+from mlx_lm.models.cache import KVCache
 from mlx_lm.models.deepseek_v3 import DeepseekV3Model
-from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 from exo.shared.models.model_cards import ModelId
 from exo.worker.engines.mlx.constants import (
-    CACHE_GROUP_SIZE,
-    KV_CACHE_BITS,
     TRUST_REMOTE_CODE,
 )
 
@@ -405,7 +402,11 @@ def apply_chat_template(
                 continue
 
             message.content = "\n".join(c.text for c in message.content).strip()
-        if message.content is None and message.thinking is None:
+        if (
+            message.content is None
+            and message.thinking is None
+            and message.tool_calls is None
+        ):
             continue
 
         # Null values are not valid when applying templates in tokenizer
@@ -462,31 +463,6 @@ class NullKVCache(KVCache):
         raise NotImplementedError("We should not be setting a NullKVCache.")
 
 
-def make_kv_cache(
-    model: Model, max_kv_size: int | None = None, keep: int = 0
-) -> list[KVCache | RotatingKVCache | QuantizedKVCache]:
-    assert hasattr(model, "layers")
-
-    # TODO: Do this for all models
-    if hasattr(model, "make_cache") and isinstance(model, GptOssModel):
-        logger.info("Using MLX LM's make cache")
-        return model.make_cache()  # type: ignore
-
-    if max_kv_size is None:
-        if KV_CACHE_BITS is None:
-            logger.info("Using default KV cache")
-            return [KVCache() for _ in model.layers]
-        else:
-            logger.info("Using quantized KV cache")
-            return [
-                QuantizedKVCache(group_size=CACHE_GROUP_SIZE, bits=KV_CACHE_BITS)
-                for _ in model.layers
-            ]
-    else:
-        logger.info(f"Using rotating KV cache with {max_kv_size=} with {keep=}")
-        return [RotatingKVCache(max_size=max_kv_size, keep=keep) for _ in model.layers]
-
-
 def mlx_force_oom(size: int = 40000) -> None:
     """
     Force an Out-Of-Memory (OOM) error in MLX by performing large tensor operations.

src/exo/worker/main.py

@@ -33,6 +33,7 @@ from exo.shared.types.events import (
 from exo.shared.types.multiaddr import Multiaddr
 from exo.shared.types.state import State
 from exo.shared.types.tasks import (
+    ChatCompletion,
     CreateRunner,
     DownloadModel,
     ImageEdits,
@@ -184,8 +185,10 @@ class Worker:
                 self.input_chunk_counts,
             )
             if task is None:
+                # Only sleep when there's nothing to do - allows rapid task dispatch
+                await anyio.sleep(0.01)
                 continue
-            logger.info(f"Worker plan: {task.__class__.__name__}")
+            logger.debug(f"Worker plan: {task.__class__.__name__}")
             assert task.task_status
             await self.event_sender.send(TaskCreated(task_id=task.task_id, task=task))
 
@@ -269,6 +272,12 @@ class Worker:
                     await self.runners[self._task_to_runner_id(task)].start_task(
                         modified_task
                     )
+                case ChatCompletion():
+                    # Don't wait for acknowledgment for batchable inference tasks
+                    # This allows multiple tasks to reach the runner for batching
+                    await self.runners[self._task_to_runner_id(task)].start_task(
+                        task, wait_for_ack=False
+                    )
                 case task:
                     await self.runners[self._task_to_runner_id(task)].start_task(task)
 

src/exo/worker/plan.py

@@ -5,6 +5,7 @@ from collections.abc import Mapping, Sequence
 from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.tasks import (
     ChatCompletion,
+    Completion,
     ConnectToGroup,
     CreateRunner,
     DownloadModel,
@@ -273,9 +274,11 @@ def _pending_tasks(
     input_chunk_buffer: Mapping[CommandId, dict[int, str]] | None = None,
 ) -> Task | None:
     for task in tasks.values():
-        # for now, just forward chat completions
+        # for now, just forward chat completions and completions
         # TODO(ciaran): do this better!
-        if not isinstance(task, (ChatCompletion, ImageGeneration, ImageEdits)):
+        if not isinstance(
+            task, (ChatCompletion, Completion, ImageGeneration, ImageEdits)
+        ):
             continue
         if task.task_status not in (TaskStatus.Pending, TaskStatus.Running):
             continue
@@ -298,9 +301,14 @@ def _pending_tasks(
             if task.task_id in runner.completed:
                 continue
 
+            # Skip tasks already sent to runner (waiting for completion)
+            if task.task_id in runner.sent:
+                continue
+
             # TODO: Check ordering aligns with MLX distributeds expectations.
 
-            if isinstance(runner.status, RunnerReady) and all(
+            # Allow sending tasks when runner is Ready OR Running (for batching)
+            if isinstance(runner.status, (RunnerReady, RunnerRunning)) and all(
                 isinstance(all_runners[global_runner_id], (RunnerReady, RunnerRunning))
                 for global_runner_id in runner.bound_instance.instance.shard_assignments.runner_to_shard
             ):

src/exo/worker/runner/batched_handler.py

@@ -0,0 +1,662 @@
+"""Batched inference handler for processing multiple ChatCompletion requests concurrently."""
+
+import time
+from collections.abc import Generator
+from dataclasses import dataclass, field
+from typing import Any, Callable, Literal
+
+import mlx.core as mx
+from mlx_lm.generate import BatchGenerator
+from mlx_lm.models.gpt_oss import Model as GptOssModel
+from mlx_lm.sample_utils import make_sampler
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
+    HarmonyEncodingName,
+    Role,
+    StreamableParser,
+    load_harmony_encoding,
+)
+
+from exo.shared.models.model_cards import ModelId
+from exo.shared.types.api import (
+    GenerationStats,
+    TopLogprobItem,
+)
+from exo.shared.types.chunks import ErrorChunk, TokenChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.events import ChunkGenerated, Event
+from exo.shared.types.memory import Memory
+from exo.shared.types.tasks import ChatCompletion
+from exo.worker.engines.mlx import Model
+from exo.worker.engines.mlx.constants import MAX_TOKENS
+from exo.worker.engines.mlx.generator.generate import extract_top_logprobs
+from exo.worker.engines.mlx.utils_mlx import apply_chat_template
+from exo.worker.runner.bootstrap import logger
+from exo.worker.runner.pipelined_generator import PipelinedGenerator, PipelinedResponse
+
+# Type alias for the finish_reason values TokenChunk accepts
+TokenFinishReason = Literal["stop", "length", "content_filter"]
+
+
+@dataclass
+class PendingRequest:
+    """A request waiting to be added to the batch."""
+
+    task: ChatCompletion
+    prompt: str
+    max_tokens: int
+    sampler: Callable[[mx.array], mx.array]
+    should_extract_logprobs: bool
+    top_k: int
+
+
+@dataclass
+class ActiveRequest:
+    """A request currently being processed in the batch."""
+
+    command_id: CommandId
+    should_extract_logprobs: bool
+    top_k: int
+    harmony_parser: Any | None = None  # StreamableParser for GPT-OSS models
+    in_thinking: bool = False  # Currently in thinking/reasoning section
+    tokens_generated: int = 0
+    reasoning_tokens: int = 0
+    prompt_tokens: int = 0
+    start_time: float = field(default_factory=time.perf_counter)
+
+
+class BatchedInferenceHandler:
+    """
+    Handles batched inference for multiple ChatCompletion requests.
+
+    Uses MLX-LM's BatchGenerator to process multiple requests concurrently,
+    improving throughput for scenarios with multiple concurrent requests.
+    """
+
+    def __init__(
+        self,
+        model: Model,
+        tokenizer: TokenizerWrapper,
+        model_id: ModelId,
+        device_rank: int,
+        world_size: int = 1,
+        max_batch_size: int = 32,
+    ):
+        self.model = model
+        self.tokenizer = tokenizer
+        self.model_id = model_id
+        self.device_rank = device_rank
+        self.world_size = world_size
+        self.max_batch_size = max_batch_size
+
+        # Model-specific thinking/reasoning detection
+        self.is_gpt_oss = isinstance(model, GptOssModel)
+        self._harmony_encoding: Any | None = None
+        if self.is_gpt_oss:
+            self._harmony_encoding = load_harmony_encoding(
+                HarmonyEncodingName.HARMONY_GPT_OSS
+            )
+            logger.info("GPT-OSS model detected, enabling harmony stream parsing")
+
+        # Detect <think></think> tokens from tokenizer (works for any model)
+        self._think_start_token: int | None = None
+        self._think_end_token: int | None = None
+        think_start: int | None = tokenizer.think_start_id  # pyright: ignore[reportAny]
+        if not self.is_gpt_oss and think_start is not None:
+            self._think_start_token = think_start
+            self._think_end_token = tokenizer.think_end_id  # pyright: ignore[reportAny]
+            logger.info(
+                f"Detected <think></think> tokens ({self._think_start_token}/{self._think_end_token}), enabling reasoning tracking"
+            )
+
+        # Pending requests waiting to be batched
+        self.pending: list[PendingRequest] = []
+
+        # Active batch generator and request tracking
+        self.batch_generator: BatchGenerator | None = None
+        self.pipelined_generator: PipelinedGenerator | None = None
+        self.uid_to_request: dict[int, ActiveRequest] = {}
+
+        # Use pipelined generator for multi-device pipeline parallelism
+        self.use_pipelined = world_size > 1
+        if self.use_pipelined:
+            logger.info(
+                f"Using PipelinedGenerator with {world_size} streams for pipeline overlap"
+            )
+
+        # EOS tokens for the model
+        self.stop_tokens: set[int] = set()
+        eos_ids: list[int] | None = getattr(tokenizer, "eos_token_ids", None)
+        if eos_ids:
+            self.stop_tokens = set(eos_ids)
+
+    @property
+    def is_active(self) -> bool:
+        """Check if there's an active batch being processed."""
+        if self.use_pipelined:
+            return (
+                self.pipelined_generator is not None
+                and self.pipelined_generator.has_active
+            )
+        return self.batch_generator is not None and len(self.uid_to_request) > 0
+
+    @property
+    def has_pending(self) -> bool:
+        """Check if there are pending requests waiting to be batched."""
+        return len(self.pending) > 0
+
+    @property
+    def current_batch_size(self) -> int:
+        """Current number of active requests in the batch."""
+        return len(self.uid_to_request)
+
+    def add_request(self, task: ChatCompletion) -> None:
+        """Add a ChatCompletion request to the pending batch."""
+        task_params = task.task_params
+
+        # Build prompt
+        prompt = apply_chat_template(self.tokenizer, task_params)
+
+        # Determine max tokens
+        max_tokens = task_params.max_tokens or MAX_TOKENS
+
+        # Create sampler for this request
+        sampler = make_sampler(
+            temp=task_params.temperature
+            if task_params.temperature is not None
+            else 0.7,
+            top_p=task_params.top_p if task_params.top_p is not None else 1.0,
+        )
+
+        # Logprobs configuration
+        should_extract_logprobs = task_params.logprobs is True
+        top_k = task_params.top_logprobs if task_params.top_logprobs is not None else 0
+
+        pending_request = PendingRequest(
+            task=task,
+            prompt=prompt,
+            max_tokens=max_tokens,
+            sampler=sampler,
+            should_extract_logprobs=should_extract_logprobs,
+            top_k=top_k,
+        )
+
+        self.pending.append(pending_request)
+
+        logger.info(
+            f"Added request to batch queue (pending={len(self.pending)}, active={self.current_batch_size})"
+        )
+
+    def flush(self) -> None:
+        """Start processing pending requests by adding them to the batch/pipelined generator."""
+        if not self.has_pending:
+            return
+
+        # Determine how many requests to flush (up to available slots)
+        available_slots = self.max_batch_size - self.current_batch_size
+        requests_to_flush = self.pending[:available_slots]
+        self.pending = self.pending[available_slots:]
+
+        # Prepare batch data - tokenize prompts
+        tokenized_prompts: list[list[int]] = []
+        max_tokens_list: list[int] = []
+        samplers: list[Callable[[mx.array], mx.array]] = []
+        prompt_token_counts: list[int] = []
+
+        for req in requests_to_flush:
+            tokens = self.tokenizer.encode(req.prompt)
+            tokenized_prompts.append(tokens)
+            max_tokens_list.append(req.max_tokens)
+            samplers.append(req.sampler)
+            prompt_token_counts.append(len(tokens))
+
+        if self.use_pipelined:
+            self._flush_pipelined(
+                requests_to_flush,
+                tokenized_prompts,
+                max_tokens_list,
+                samplers,
+                prompt_token_counts,
+            )
+        else:
+            self._flush_batch(
+                requests_to_flush,
+                tokenized_prompts,
+                max_tokens_list,
+                samplers,
+                prompt_token_counts,
+            )
+
+    def _flush_pipelined(
+        self,
+        requests_to_flush: list[PendingRequest],
+        tokenized_prompts: list[list[int]],
+        max_tokens_list: list[int],
+        samplers: list[Callable[[mx.array], mx.array]],
+        prompt_token_counts: list[int],
+    ) -> None:
+        """Flush using PipelinedGenerator (multi-stream pipeline overlap)."""
+        if self.pipelined_generator is None:
+            logger.info(
+                f"Creating PipelinedGenerator for {len(requests_to_flush)} requests ({self.world_size} streams)"
+            )
+            mx.reset_peak_memory()
+            self.pipelined_generator = PipelinedGenerator(
+                model=self.model,
+                world_size=self.world_size,
+                stop_tokens=self.stop_tokens if self.stop_tokens else None,
+                max_tokens=MAX_TOKENS,
+            )
+        else:
+            logger.info(
+                f"Adding {len(requests_to_flush)} requests to PipelinedGenerator"
+            )
+
+        uids = self.pipelined_generator.insert(
+            prompts=tokenized_prompts,
+            max_tokens=max_tokens_list,
+            samplers=samplers,
+        )
+
+        for uid, req, prompt_tokens, tokens in zip(
+            uids, requests_to_flush, prompt_token_counts, tokenized_prompts, strict=True
+        ):
+            parser = None
+            if self.is_gpt_oss and self._harmony_encoding is not None:
+                parser = StreamableParser(self._harmony_encoding, role=Role.ASSISTANT)  # pyright: ignore[reportAny]
+            # Check if prompt contains <think> token - if so, model is already in thinking mode
+            starts_in_thinking = (
+                self._think_start_token is not None
+                and self._think_start_token in tokens
+            )
+            self.uid_to_request[uid] = ActiveRequest(
+                command_id=req.task.command_id,
+                should_extract_logprobs=req.should_extract_logprobs,
+                top_k=req.top_k,
+                prompt_tokens=prompt_tokens,
+                harmony_parser=parser,
+                in_thinking=starts_in_thinking,
+            )
+
+        logger.info(
+            f"Flushed {len(requests_to_flush)} requests into pipelined generator (active={self.pipelined_generator.active_count}, uids={list(self.uid_to_request.keys())})"
+        )
+
+    def _flush_batch(
+        self,
+        requests_to_flush: list[PendingRequest],
+        tokenized_prompts: list[list[int]],
+        max_tokens_list: list[int],
+        samplers: list[Callable[[mx.array], mx.array]],
+        prompt_token_counts: list[int],
+    ) -> None:
+        """Flush using BatchGenerator (single-stream, for non-pipeline instances)."""
+        if self.batch_generator is None:
+            logger.info(
+                f"Creating new BatchGenerator for {len(requests_to_flush)} requests"
+            )
+            mx.reset_peak_memory()
+            self.batch_generator = BatchGenerator(
+                model=self.model,
+                max_tokens=MAX_TOKENS,
+                stop_tokens=self.stop_tokens if self.stop_tokens else None,
+                prefill_batch_size=1,
+            )
+        else:
+            logger.info(
+                f"Adding {len(requests_to_flush)} requests to existing BatchGenerator"
+            )
+
+        # Insert into batch generator
+        uids: list[int] = self.batch_generator.insert(  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            prompts=tokenized_prompts,
+            max_tokens=max_tokens_list,
+            samplers=samplers,  # pyright: ignore[reportCallIssue]
+        )
+
+        for uid, req, prompt_tokens, tokens in zip(
+            uids, requests_to_flush, prompt_token_counts, tokenized_prompts, strict=True
+        ):  # pyright: ignore[reportUnknownArgumentType]
+            parser = None
+            if self.is_gpt_oss and self._harmony_encoding is not None:
+                parser = StreamableParser(self._harmony_encoding, role=Role.ASSISTANT)  # pyright: ignore[reportAny]
+            # Check if prompt contains <think> token - if so, model is already in thinking mode
+            starts_in_thinking = (
+                self._think_start_token is not None
+                and self._think_start_token in tokens
+            )
+            self.uid_to_request[uid] = ActiveRequest(
+                command_id=req.task.command_id,
+                should_extract_logprobs=req.should_extract_logprobs,
+                top_k=req.top_k,
+                prompt_tokens=prompt_tokens,
+                harmony_parser=parser,
+                in_thinking=starts_in_thinking,
+            )
+
+        logger.info(
+            f"Flushed {len(requests_to_flush)} requests into batch (active={self.current_batch_size}, uids={list(self.uid_to_request.keys())})"
+        )
+
+    def step(self) -> Generator[Event, None, None]:
+        """
+        Process one generation step and yield ChunkGenerated events.
+
+        Returns a generator of events for completed tokens across all active requests.
+        """
+        if self.use_pipelined:
+            yield from self._step_pipelined()
+            return
+
+        if self.batch_generator is None or not self.uid_to_request:
+            return
+
+        # Get next tokens for all active requests
+        # BatchGenerator.next() returns list of Response objects
+        logger.debug(
+            f"BatchGenerator.next() called (active_uids={list(self.uid_to_request.keys())})"
+        )
+        responses: list[Any] = self.batch_generator.next()  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+        logger.debug(f"BatchGenerator.next() returned {len(responses)} responses")  # pyright: ignore[reportUnknownArgumentType]
+
+        completed_uids: list[int] = []
+
+        for response in responses:  # pyright: ignore[reportUnknownVariableType]
+            uid: int = response.uid  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            if uid not in self.uid_to_request:
+                logger.warning(f"Received response for unknown uid: {uid}")
+                continue
+
+            active_request = self.uid_to_request[uid]
+            active_request.tokens_generated += 1
+
+            # Extract response fields with explicit typing
+            resp_token: int = response.token  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            resp_finish_reason: str | None = response.finish_reason  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+            resp_logprobs: mx.array = response.logprobs  # pyright: ignore[reportUnknownMemberType,reportUnknownVariableType]
+
+            # Only emit events from device_rank 0
+            if self.device_rank != 0:
+                if resp_finish_reason is not None:
+                    completed_uids.append(uid)  # pyright: ignore[reportUnknownArgumentType]
+                continue
+
+            # Decode token to text
+            # Skip emitting EOS token text (e.g., <|eot_id|>)
+            if resp_token in self.stop_tokens:
+                token_text = ""
+            else:
+                token_text = self.tokenizer.decode([resp_token])
+
+            # Handle thinking/reasoning token tracking
+            if active_request.harmony_parser is not None:
+                # GPT-OSS: Use harmony parser for channel-based thinking detection
+                parser = active_request.harmony_parser  # pyright: ignore[reportAny]
+                parser.process(resp_token)  # pyright: ignore[reportAny]
+                delta: str | None = parser.last_content_delta  # pyright: ignore[reportAny]
+                channel: str = parser.current_channel  # pyright: ignore[reportAny]
+
+                # Track reasoning tokens (analysis channel = thinking)
+                if channel == "analysis":
+                    active_request.reasoning_tokens += 1
+
+                # Handle thinking tag transitions
+                prefix = ""
+                if channel == "analysis" and not active_request.in_thinking:
+                    active_request.in_thinking = True
+                    prefix = "<think>"
+                elif channel != "analysis" and active_request.in_thinking:
+                    active_request.in_thinking = False
+                    prefix = "</think>"
+
+                if resp_finish_reason is not None and active_request.in_thinking:
+                    # Close thinking tag on finish
+                    prefix = "</think>"
+                    active_request.in_thinking = False
+
+                effective_delta = delta or ""
+                token_text = (
+                    prefix + effective_delta if (prefix or effective_delta) else ""
+                )
+                # Skip empty tokens (channel markers with no content delta)
+                if not token_text and resp_finish_reason is None:
+                    continue
+            elif self._think_start_token is not None:
+                # MiniMax: Track <think>/</ think> tokens directly
+                if resp_token == self._think_start_token:
+                    active_request.in_thinking = True
+                elif resp_token == self._think_end_token:
+                    active_request.in_thinking = False
+                elif active_request.in_thinking:
+                    active_request.reasoning_tokens += 1
+
+            # Extract logprobs if requested
+            logprob: float | None = None
+            top_logprobs: list[TopLogprobItem] | None = None
+            if active_request.should_extract_logprobs:
+                logprob, top_logprobs = extract_top_logprobs(
+                    logprobs_array=resp_logprobs,  # pyright: ignore[reportUnknownArgumentType]
+                    selected_token=resp_token,  # pyright: ignore[reportUnknownArgumentType]
+                    tokenizer=self.tokenizer,
+                    top_k=active_request.top_k,
+                )
+
+            # Build stats for final token
+            stats: GenerationStats | None = None
+            finish_reason: TokenFinishReason | None = None
+            if resp_finish_reason is not None:
+                elapsed_time = time.perf_counter() - active_request.start_time
+                prompt_tps = active_request.prompt_tokens / max(elapsed_time, 0.001)
+                generation_tps = active_request.tokens_generated / max(
+                    elapsed_time, 0.001
+                )
+
+                # Get peak memory
+                peak_memory_bytes = 0
+                if mx.metal.is_available():
+                    peak_memory_bytes = mx.metal.get_peak_memory()
+
+                stats = GenerationStats(
+                    prompt_tps=prompt_tps,
+                    generation_tps=generation_tps,
+                    prompt_tokens=active_request.prompt_tokens,
+                    generation_tokens=active_request.tokens_generated,
+                    reasoning_tokens=active_request.reasoning_tokens,
+                    peak_memory_usage=Memory.from_bytes(peak_memory_bytes),
+                )
+
+                # Map finish reason to the narrower type TokenChunk expects
+                if resp_finish_reason == "stop":
+                    finish_reason = "stop"
+                elif resp_finish_reason == "length":
+                    finish_reason = "length"
+                elif resp_finish_reason == "content_filter":
+                    finish_reason = "content_filter"
+                else:
+                    # Unknown finish reasons default to "stop"
+                    logger.warning(
+                        f"Unknown finish_reason: {resp_finish_reason}, mapping to 'stop'"
+                    )
+                    finish_reason = "stop"
+
+                completed_uids.append(uid)  # pyright: ignore[reportUnknownArgumentType]
+
+            yield ChunkGenerated(
+                command_id=active_request.command_id,
+                chunk=TokenChunk(
+                    model=self.model_id,
+                    text=token_text,
+                    token_id=resp_token,  # pyright: ignore[reportUnknownArgumentType]
+                    logprob=logprob,
+                    top_logprobs=top_logprobs,
+                    finish_reason=finish_reason,
+                    stats=stats,
+                ),
+            )
+
+        # Clean up completed requests
+        for uid in completed_uids:
+            del self.uid_to_request[uid]
+
+    def _step_pipelined(self) -> Generator[Event, None, None]:
+        """Process one generation step using the multi-stream PipelinedGenerator."""
+        if self.pipelined_generator is None or not self.uid_to_request:
+            return
+
+        logger.debug(
+            f"PipelinedGenerator.next() called (active={self.pipelined_generator.active_count})"
+        )
+        responses: list[PipelinedResponse] = self.pipelined_generator.next()
+        logger.debug(f"PipelinedGenerator.next() returned {len(responses)} responses")
+
+        completed_uids: list[int] = []
+
+        for response in responses:
+            uid = response.uid
+            if uid not in self.uid_to_request:
+                logger.warning(f"Received response for unknown uid: {uid}")
+                continue
+
+            active_request = self.uid_to_request[uid]
+            active_request.tokens_generated += 1
+
+            resp_token: int = response.token
+            resp_finish_reason: str | None = response.finish_reason
+            resp_logprobs: mx.array = response.logprobs
+
+            # Only emit events from device_rank 0
+            if self.device_rank != 0:
+                if resp_finish_reason is not None:
+                    completed_uids.append(uid)
+                continue
+
+            # Decode token to text
+            # Skip emitting EOS token text (e.g., <|eot_id|>)
+            if resp_token in self.stop_tokens:
+                token_text = ""
+            else:
+                token_text = self.tokenizer.decode([resp_token])
+
+            # Handle thinking/reasoning token tracking
+            if active_request.harmony_parser is not None:
+                # GPT-OSS: Use harmony parser for channel-based thinking detection
+                parser = active_request.harmony_parser  # pyright: ignore[reportAny]
+                parser.process(resp_token)  # pyright: ignore[reportAny]
+                delta: str | None = parser.last_content_delta  # pyright: ignore[reportAny]
+                channel: str = parser.current_channel  # pyright: ignore[reportAny]
+
+                if channel == "analysis":
+                    active_request.reasoning_tokens += 1
+
+                prefix = ""
+                if channel == "analysis" and not active_request.in_thinking:
+                    active_request.in_thinking = True
+                    prefix = "<think>"
+                elif channel != "analysis" and active_request.in_thinking:
+                    active_request.in_thinking = False
+                    prefix = "</think>"
+
+                if resp_finish_reason is not None and active_request.in_thinking:
+                    prefix = "</think>"
+                    active_request.in_thinking = False
+
+                effective_delta = delta or ""
+                token_text = (
+                    prefix + effective_delta if (prefix or effective_delta) else ""
+                )
+                if not token_text and resp_finish_reason is None:
+                    continue
+            elif self._think_start_token is not None:
+                # MiniMax: Track <think>/</think> tokens directly
+                if resp_token == self._think_start_token:
+                    active_request.in_thinking = True
+                elif resp_token == self._think_end_token:
+                    active_request.in_thinking = False
+                elif active_request.in_thinking:
+                    active_request.reasoning_tokens += 1
+
+            # Extract logprobs if requested
+            logprob: float | None = None
+            top_logprobs: list[TopLogprobItem] | None = None
+            if active_request.should_extract_logprobs:
+                logprob, top_logprobs = extract_top_logprobs(
+                    logprobs_array=resp_logprobs,
+                    selected_token=resp_token,
+                    tokenizer=self.tokenizer,
+                    top_k=active_request.top_k,
+                )
+
+            # Build stats for final token
+            stats: GenerationStats | None = None
+            finish_reason: TokenFinishReason | None = None
+            if resp_finish_reason is not None:
+                elapsed_time = time.perf_counter() - active_request.start_time
+                prompt_tps = active_request.prompt_tokens / max(elapsed_time, 0.001)
+                generation_tps = active_request.tokens_generated / max(
+                    elapsed_time, 0.001
+                )
+
+                peak_memory_bytes = 0
+                if mx.metal.is_available():
+                    peak_memory_bytes = mx.metal.get_peak_memory()
+
+                stats = GenerationStats(
+                    prompt_tps=prompt_tps,
+                    generation_tps=generation_tps,
+                    prompt_tokens=active_request.prompt_tokens,
+                    generation_tokens=active_request.tokens_generated,
+                    reasoning_tokens=active_request.reasoning_tokens,
+                    peak_memory_usage=Memory.from_bytes(peak_memory_bytes),
+                )
+
+                if resp_finish_reason == "stop":
+                    finish_reason = "stop"
+                elif resp_finish_reason == "length":
+                    finish_reason = "length"
+                else:
+                    finish_reason = "stop"
+
+                completed_uids.append(uid)
+
+            yield ChunkGenerated(
+                command_id=active_request.command_id,
+                chunk=TokenChunk(
+                    model=self.model_id,
+                    text=token_text,
+                    token_id=resp_token,
+                    logprob=logprob,
+                    top_logprobs=top_logprobs,
+                    finish_reason=finish_reason,
+                    stats=stats,
+                ),
+            )
+
+        for uid in completed_uids:
+            del self.uid_to_request[uid]
+
+    def emit_error(self, command_id: CommandId, error_message: str) -> Event:
+        """Create an error event for a failed request."""
+        return ChunkGenerated(
+            command_id=command_id,
+            chunk=ErrorChunk(
+                model=self.model_id,
+                finish_reason="error",
+                error_message=error_message,
+            ),
+        )
+
+    def _close_generator(self) -> None:
+        """Close and clean up the batch/pipelined generator."""
+        if self.batch_generator is not None:
+            self.batch_generator.close()  # pyright: ignore[reportUnknownMemberType,reportAttributeAccessIssue]
+            self.batch_generator = None
+        if self.pipelined_generator is not None:
+            self.pipelined_generator.close()
+            self.pipelined_generator = None
+        self.uid_to_request.clear()
+        logger.info("Generator closed")
+
+    def close(self) -> None:
+        """Close the handler and clean up resources."""
+        self._close_generator()
+        self.pending.clear()

src/exo/worker/runner/batched_scoring_handler.py

@@ -0,0 +1,200 @@
+"""Batched scoring handler for processing multiple Completion requests concurrently."""
+
+import time
+from dataclasses import dataclass, field
+
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+
+from exo.shared.models.model_cards import ModelId
+from exo.shared.types.api import TopLogprobItem
+from exo.shared.types.chunks import CompletionChunk, ErrorChunk
+from exo.shared.types.events import ChunkGenerated, Event
+from exo.shared.types.tasks import Completion
+from exo.worker.engines.mlx import Model
+from exo.worker.engines.mlx.generator.generate import score_tokens_batched
+from exo.worker.runner.bootstrap import logger
+
+
+@dataclass
+class PendingScoringRequest:
+    """A scoring request waiting to be batched."""
+
+    task: Completion
+    tokens: list[int]
+    prompt_text: str
+    top_k: int | None
+    echo: bool
+
+
+@dataclass
+class BatchedScoringHandler:
+    """
+    Handles batched scoring for multiple Completion requests.
+
+    Collects multiple scoring requests and processes them in a single
+    batched forward pass for improved throughput.
+    """
+
+    model: Model
+    tokenizer: TokenizerWrapper
+    model_id: ModelId
+    device_rank: int
+    max_batch_size: int = 32
+    batch_timeout_ms: int = 10
+
+    pending: list[PendingScoringRequest] = field(default_factory=list)
+    pending_start_time: float | None = None
+
+    @property
+    def has_pending(self) -> bool:
+        """Check if there are pending requests."""
+        return len(self.pending) > 0
+
+    def add_request(
+        self,
+        task: Completion,
+        tokens: list[int],
+        prompt_text: str,
+    ) -> None:
+        """Add a Completion request to the pending batch."""
+        task_params = task.task_params
+        top_k = task_params.logprobs
+
+        self.pending.append(
+            PendingScoringRequest(
+                task=task,
+                tokens=tokens,
+                prompt_text=prompt_text,
+                top_k=top_k,
+                echo=task_params.echo,
+            )
+        )
+
+        if self.pending_start_time is None:
+            self.pending_start_time = time.perf_counter()
+
+        logger.debug(f"Added scoring request to batch (pending={len(self.pending)})")
+
+    def should_flush(self) -> bool:
+        """Check if the batch should be flushed."""
+        if not self.has_pending:
+            return False
+
+        # Flush if batch is full
+        if len(self.pending) >= self.max_batch_size:
+            return True
+
+        # Flush if timeout reached
+        if self.pending_start_time is not None:
+            elapsed_ms = (time.perf_counter() - self.pending_start_time) * 1000
+            if elapsed_ms >= self.batch_timeout_ms:
+                return True
+
+        return False
+
+    def flush(self) -> list[Event]:
+        """Process all pending requests and return events."""
+        if not self.has_pending:
+            return []
+
+        requests = self.pending
+        self.pending = []
+        self.pending_start_time = None
+
+        logger.info(f"Processing batch of {len(requests)} scoring requests")
+
+        # Collect all token sequences
+        token_sequences = [req.tokens for req in requests]
+
+        # Get common top_k (use first request's top_k, they should all be the same)
+        top_k = requests[0].top_k if requests else None
+
+        try:
+            # Run batched scoring
+            all_results = score_tokens_batched(
+                model=self.model,
+                tokenizer=self.tokenizer,
+                token_sequences=token_sequences,
+                top_k=top_k,
+            )
+
+            # Generate events for each request
+            events: list[Event] = []
+            for req, logprob_results in zip(requests, all_results, strict=True):
+                if self.device_rank != 0:
+                    continue
+
+                event = self._build_completion_event(req, logprob_results)
+                events.append(event)
+
+            logger.info(f"Batch scoring complete ({len(events)} events)")
+            return events
+
+        except Exception as e:
+            # Return error events for all requests
+            logger.error(f"Batch scoring failed: {e}")
+            events = []
+            for req in requests:
+                if self.device_rank == 0:
+                    events.append(
+                        ChunkGenerated(
+                            command_id=req.task.command_id,
+                            chunk=ErrorChunk(
+                                model=self.model_id,
+                                finish_reason="error",
+                                error_message=str(e),
+                            ),
+                        )
+                    )
+            return events
+
+    def _build_completion_event(
+        self,
+        req: PendingScoringRequest,
+        logprob_results: list[tuple[float, list[TopLogprobItem]]],
+    ) -> Event:
+        """Build a ChunkGenerated event for a completed scoring request."""
+        tokens = req.tokens
+        tokenizer = self.tokenizer
+
+        # Build response in completions format
+        token_strings: list[str] = []
+        token_logprobs: list[float | None] = []
+        top_logprobs: list[dict[str, float]] = []
+        text_offset: list[int] = []
+
+        offset = 0
+        for i, token_id in enumerate(tokens):
+            token_str = tokenizer.decode([token_id])
+            token_strings.append(token_str)
+
+            if i < len(logprob_results):
+                logprob, top_items = logprob_results[i]
+                # First token has no logprob (None in OpenAI format)
+                token_logprobs.append(logprob if i > 0 else None)
+                top_lp_dict = {item.token: item.logprob for item in top_items}
+                top_logprobs.append(top_lp_dict)
+            else:
+                token_logprobs.append(None)
+                top_logprobs.append({})
+
+            text_offset.append(offset)
+            offset += len(token_str)
+
+        return ChunkGenerated(
+            command_id=req.task.command_id,
+            chunk=CompletionChunk(
+                model=self.model_id,
+                text=req.prompt_text if req.echo else "",
+                tokens=token_strings,
+                token_logprobs=token_logprobs,
+                top_logprobs=top_logprobs,
+                text_offset=text_offset,
+                finish_reason="stop",
+            ),
+        )
+
+    def close(self) -> None:
+        """Clean up resources."""
+        self.pending.clear()
+        self.pending_start_time = None

src/exo/worker/runner/pipelined_generator.py

@@ -0,0 +1,334 @@
+"""Multi-stream pipelined batch generator for pipeline-parallel inference.
+
+When a model is split across N ranks (pipeline parallelism), each rank's GPU is idle
+for (N-1)/N of each step while waiting for other ranks to compute their layers.
+
+This module fills the pipeline bubble by splitting sequences into N micro-batch groups
+and processing each group on a different MLX stream. The GPU can overlap one stream's
+network communication (send/recv/all_gather) with another stream's compute.
+"""
+
+# pyright: reportUnknownMemberType=false, reportUnknownVariableType=false
+# pyright: reportUnknownArgumentType=false, reportAny=false
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx_lm.models.cache import make_prompt_cache
+
+
+@dataclass
+class MicroBatch:
+    """State for one micro-batch group of sequences."""
+
+    uids: list[int]
+    y: mx.array  # Last sampled tokens [batch]
+    logprobs: list[mx.array]  # Logprobs for each sequence
+    max_tokens: list[int]
+    num_tokens: list[int]
+    cache: list[Any]  # KV cache (list of layer caches)
+    samplers: list[Callable[[mx.array], mx.array]]
+    tokens: list[mx.array]  # All tokens generated so far per sequence
+
+    def __len__(self) -> int:
+        return len(self.uids)
+
+
+@dataclass
+class PipelinedResponse:
+    """Response from one generation step."""
+
+    uid: int
+    token: int
+    logprobs: mx.array
+    finish_reason: str | None
+    cache: list[Any] | None = None
+
+
+@dataclass
+class PendingPrompt:
+    """A prompt waiting to be prefilled."""
+
+    uid: int
+    tokens: list[int]
+    max_tokens: int
+    sampler: Callable[[mx.array], mx.array]
+
+
+class PipelinedGenerator:
+    """
+    Multi-stream batch generator that fills pipeline bubbles.
+
+    Splits active sequences into `world_size` micro-batch groups, each processed
+    on its own MLX stream. During mx.eval(), the GPU overlaps network operations
+    on one stream with compute on another.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        world_size: int,
+        stop_tokens: set[int] | None = None,
+        max_tokens: int = 4096,
+    ):
+        self.model = model
+        self.world_size = world_size
+        self.stop_tokens = stop_tokens or set()
+        self.max_tokens_default = max_tokens
+
+        # Create one stream per pipeline stage
+        self.streams = [mx.new_stream(mx.default_device()) for _ in range(world_size)]
+
+        # Micro-batch groups (one per stream)
+        self.micro_batches: list[MicroBatch | None] = [None] * world_size
+
+        # Pending prompts to be inserted
+        self.pending_prompts: list[PendingPrompt] = []
+
+        # UID counter
+        self._next_uid = 0
+
+    @property
+    def active_count(self) -> int:
+        """Total number of active sequences across all micro-batches."""
+        return sum(len(mb) for mb in self.micro_batches if mb is not None)
+
+    @property
+    def has_active(self) -> bool:
+        return self.active_count > 0 or len(self.pending_prompts) > 0
+
+    def insert(
+        self,
+        prompts: list[list[int]],
+        max_tokens: list[int],
+        samplers: list[Callable[[mx.array], mx.array]],
+    ) -> list[int]:
+        """Queue prompts for processing. Returns assigned UIDs."""
+        uids: list[int] = []
+        for prompt, mt, sampler in zip(prompts, max_tokens, samplers, strict=True):
+            uid = self._next_uid
+            self._next_uid += 1
+            self.pending_prompts.append(
+                PendingPrompt(uid=uid, tokens=prompt, max_tokens=mt, sampler=sampler)
+            )
+            uids.append(uid)
+        return uids
+
+    def _prefill_group(self, group_idx: int, prompts: list[PendingPrompt]) -> None:
+        """Prefill a group of prompts and create a MicroBatch."""
+        if not prompts:
+            return
+
+        stream = self.streams[group_idx]
+
+        with mx.stream(stream):
+            # Create per-sequence caches
+            caches = [make_prompt_cache(self.model) for _ in prompts]
+
+            # Tokenize and prefill each sequence
+            all_y: list[mx.array] = []
+            all_logprobs: list[mx.array] = []
+            all_samplers: list[Callable[[mx.array], mx.array]] = []
+            all_tokens: list[mx.array] = []
+
+            for prompt_info, cache in zip(prompts, caches, strict=True):
+                tokens = mx.array(prompt_info.tokens)
+                # Run prefill (process all tokens except last)
+                if len(prompt_info.tokens) > 1:
+                    self.model(tokens[:-1][None, :], cache=cache)
+                    mx.eval([c.state for c in cache])
+
+                # Process last token to get first generation logits
+                last_token = tokens[-1:][None, :]
+                logits = self.model(last_token, cache=cache)
+                logits = logits[:, -1, :]
+                logprobs = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
+                sampled = prompt_info.sampler(logprobs)
+
+                all_y.append(sampled.squeeze(0))
+                all_logprobs.append(logprobs.squeeze(0))
+                all_samplers.append(prompt_info.sampler)
+                all_tokens.append(tokens)
+
+            mx.eval(*all_y, *all_logprobs)
+
+        # Create micro-batch
+        batch = MicroBatch(
+            uids=[p.uid for p in prompts],
+            y=mx.stack(all_y),
+            logprobs=all_logprobs,
+            max_tokens=[p.max_tokens for p in prompts],
+            num_tokens=[0] * len(prompts),
+            cache=caches,
+            samplers=all_samplers,
+            tokens=all_tokens,
+        )
+
+        if self.micro_batches[group_idx] is None:
+            self.micro_batches[group_idx] = batch
+        else:
+            # Extend existing micro-batch (would need cache merging - for now replace)
+            self.micro_batches[group_idx] = batch
+
+    def _prefill_pending(self) -> None:
+        """Distribute pending prompts across micro-batch groups and prefill."""
+        if not self.pending_prompts:
+            return
+
+        # Distribute round-robin across groups
+        groups: list[list[PendingPrompt]] = [[] for _ in range(self.world_size)]
+        for i, prompt in enumerate(self.pending_prompts):
+            groups[i % self.world_size].append(prompt)
+        self.pending_prompts.clear()
+
+        for group_idx, group_prompts in enumerate(groups):
+            if group_prompts:
+                self._prefill_group(group_idx, group_prompts)
+
+    def _step_all(self) -> None:
+        """
+        Run one generation step across all micro-batch groups on different streams.
+
+        This is where pipeline overlap happens: each group's model forward pass
+        runs on its own stream, and mx.eval() allows the GPU to overlap network
+        ops (send/recv/all_gather) from one stream with compute from another.
+
+        Each sequence is processed individually with its own KV cache, but all
+        lazy graphs across streams are evaluated together for GPU overlap.
+        """
+        # Build computation graphs on each stream (lazy, no evaluation yet)
+        # Each micro-batch group processes its sequences on its own stream.
+        all_sampled: list[mx.array] = []
+        all_logprobs: list[mx.array] = []
+        # Track which (group_idx, seq_idx) each result corresponds to
+        result_map: list[tuple[int, int]] = []
+
+        for i, mb in enumerate(self.micro_batches):
+            if mb is None or len(mb) == 0:
+                continue
+
+            with mx.stream(self.streams[i]):
+                for e in range(len(mb)):
+                    # Process each sequence individually with its own cache
+                    input_token = mb.y[e : e + 1][None, :]  # [1, 1]
+
+                    # Forward pass (lazy graph construction)
+                    # For pipeline models, this includes send/recv/all_gather ops
+                    logits = self.model(input_token, cache=mb.cache[e])
+                    logits = logits[:, -1, :]  # [1, vocab]
+
+                    # Compute logprobs
+                    logprobs = logits - mx.logsumexp(logits, axis=-1, keepdims=True)
+
+                    # Sample
+                    sampled = mb.samplers[e](logprobs)
+
+                    all_sampled.append(sampled.squeeze(0))
+                    all_logprobs.append(logprobs.squeeze(0))
+                    result_map.append((i, e))
+
+        if not result_map:
+            return
+
+        # Evaluate ALL streams together - this is where overlap happens!
+        # The GPU can execute stream0's all_gather while computing stream1's layers.
+        mx.eval(*all_sampled, *all_logprobs)
+
+        # Update micro-batch states with results
+        # Group results by micro-batch for efficient update
+        group_results: dict[int, list[int]] = {}
+        for idx, (group_idx, _seq_idx) in enumerate(result_map):
+            group_results.setdefault(group_idx, []).append(idx)
+
+        for group_idx, result_indices in group_results.items():
+            mb = self.micro_batches[group_idx]
+            assert mb is not None
+            group_sampled = [all_sampled[idx] for idx in result_indices]
+            group_logprobs = [all_logprobs[idx] for idx in result_indices]
+            mb.y = mx.stack(group_sampled)
+            mb.logprobs = group_logprobs
+            for e, idx in enumerate(result_indices):
+                mb.tokens[e] = mx.concatenate([mb.tokens[e], all_sampled[idx][None]])
+
+    def next(self) -> list[PipelinedResponse]:
+        """
+        Run one generation step and return responses.
+
+        Returns a PipelinedResponse for each active sequence (across all groups).
+        Finished sequences are removed from their micro-batch.
+        """
+        # Prefill any pending prompts first
+        self._prefill_pending()
+
+        if not self.has_active:
+            return []
+
+        # Run the multi-stream forward pass
+        self._step_all()
+
+        # Collect responses and filter completed sequences
+        responses: list[PipelinedResponse] = []
+
+        for group_idx, mb in enumerate(self.micro_batches):
+            if mb is None or len(mb) == 0:
+                continue
+
+            keep_idx: list[int] = []
+            end_idx: list[int] = []
+
+            for e in range(len(mb)):
+                token = int(mb.y[e].item())
+                uid = mb.uids[e]
+                num_tok = mb.num_tokens[e] + 1
+                max_tok = mb.max_tokens[e]
+                mb.num_tokens[e] = num_tok
+
+                if token in self.stop_tokens:
+                    finish_reason = "stop"
+                    end_idx.append(e)
+                elif num_tok >= max_tok:
+                    finish_reason = "length"
+                    end_idx.append(e)
+                else:
+                    finish_reason = None
+                    keep_idx.append(e)
+
+                responses.append(
+                    PipelinedResponse(
+                        uid=uid,
+                        token=token,
+                        logprobs=mb.logprobs[e],
+                        finish_reason=finish_reason,
+                    )
+                )
+
+            # Remove finished sequences
+            if end_idx:
+                if keep_idx:
+                    # Filter the micro-batch to keep only active sequences
+                    mb.uids = [mb.uids[i] for i in keep_idx]
+                    mb.y = mb.y[mx.array(keep_idx)]
+                    mb.logprobs = [mb.logprobs[i] for i in keep_idx]
+                    mb.max_tokens = [mb.max_tokens[i] for i in keep_idx]
+                    mb.num_tokens = [mb.num_tokens[i] for i in keep_idx]
+                    mb.samplers = [mb.samplers[i] for i in keep_idx]
+                    mb.tokens = [mb.tokens[i] for i in keep_idx]
+                    # Cache filtering: trim batch dimension
+                    for c in mb.cache:
+                        if hasattr(c, "keys") and c.keys is not None:
+                            c.keys = c.keys[mx.array(keep_idx)]
+                            c.values = c.values[mx.array(keep_idx)]
+                else:
+                    self.micro_batches[group_idx] = None
+
+        return responses
+
+    def close(self) -> None:
+        """Clean up resources."""
+        self.micro_batches = [None] * self.world_size
+        self.pending_prompts.clear()

src/exo/worker/runner/runner.py

@@ -6,6 +6,7 @@ from functools import cache
 from typing import Any, Callable, Literal
 
 import mlx.core as mx
+from anyio import EndOfStream, WouldBlock
 from mlx_lm.models.gpt_oss import Model as GptOssModel
 from mlx_lm.tokenizer_utils import TokenizerWrapper
 from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
@@ -61,7 +62,7 @@ from exo.shared.types.worker.runners import (
     RunnerStatus,
     RunnerWarmingUp,
 )
-from exo.shared.types.worker.shards import ShardMetadata
+from exo.shared.types.worker.shards import ShardMetadata, TensorShardMetadata
 from exo.utils.channels import MpReceiver, MpSender
 from exo.worker.engines.image import (
     DistributedImageModel,
@@ -70,7 +71,11 @@ from exo.worker.engines.image import (
     warmup_image_generator,
 )
 from exo.worker.engines.mlx import Model
-from exo.worker.engines.mlx.generator.generate import mlx_generate, warmup_inference
+from exo.worker.engines.mlx.cache import KVPrefixCache
+from exo.worker.engines.mlx.generator.generate import (
+    mlx_generate,
+    warmup_inference,
+)
 from exo.worker.engines.mlx.utils_mlx import (
     apply_chat_template,
     detect_thinking_prompt_suffix,
@@ -78,8 +83,128 @@ from exo.worker.engines.mlx.utils_mlx import (
     load_mlx_items,
     mlx_force_oom,
 )
+from exo.worker.runner.batched_handler import BatchedInferenceHandler
 from exo.worker.runner.bootstrap import logger
 
+# Batching configuration
+BATCH_ENABLED = True
+BATCH_MAX_SIZE = 32
+
+
+def _should_use_serial_processing(
+    task: ChatCompletion,
+    tokenizer: TokenizerWrapper,
+    model: Model,
+    model_id: ModelId,
+) -> bool:
+    """
+    Determine if a ChatCompletion task requires serial processing.
+
+    Currently always returns False - batch mode handles all cases.
+    Post-processing (GPT-OSS, thinking models, tool calls) can be applied
+    per-request to the individual streams from the batch generator.
+    """
+    # All tasks can use batch mode - post-processing is per-request
+    return False
+
+
+def _process_serial_chat_completion(
+    task: ChatCompletion,
+    model: Model,
+    tokenizer: TokenizerWrapper,
+    shard_metadata: ShardMetadata,
+    event_sender: MpSender[Event],
+) -> None:
+    """Process a ChatCompletion task serially (original implementation)."""
+    task_params = task.task_params
+    command_id = task.command_id
+    device_rank = shard_metadata.device_rank
+
+    if task_params.messages[0].content is not None:
+        _check_for_debug_prompts(task_params.messages[0].content)
+
+    # Build prompt once - used for both generation and thinking detection
+    prompt = apply_chat_template(tokenizer, task_params)
+
+    # Generate responses using the actual MLX generation
+    mlx_generator = mlx_generate(
+        model=model,
+        tokenizer=tokenizer,
+        task=task_params,
+        prompt=prompt,
+    )
+
+    # GPT-OSS specific parsing to match other model formats.
+    if isinstance(model, GptOssModel):
+        mlx_generator = parse_gpt_oss(mlx_generator)
+
+    # For other thinking models (GLM, etc.), check if we need to
+    # prepend the thinking tag that was consumed by the chat template
+    if detect_thinking_prompt_suffix(prompt, tokenizer):
+        mlx_generator = parse_thinking_models(mlx_generator, tokenizer)
+
+    # Kimi-K2 has tool call sections - we don't care about them
+    if "kimi" in shard_metadata.model_card.model_id.lower():
+        mlx_generator = filter_kimi_tokens(mlx_generator)
+        patch_kimi_tokenizer(tokenizer)
+
+    if tokenizer.has_tool_calling:
+        assert tokenizer.tool_call_start
+        assert tokenizer.tool_call_end
+        assert tokenizer.tool_parser  # pyright: ignore[reportAny]
+        mlx_generator = parse_tool_calls(
+            mlx_generator,
+            tokenizer.tool_call_start,
+            tokenizer.tool_call_end,
+            tokenizer.tool_parser,  # pyright: ignore[reportAny]
+        )
+
+    for response in mlx_generator:
+        match response:
+            case GenerationResponse():
+                if device_rank == 0 and response.finish_reason == "error":
+                    event_sender.send(
+                        ChunkGenerated(
+                            command_id=command_id,
+                            chunk=ErrorChunk(
+                                error_message=response.text,
+                                model=shard_metadata.model_card.model_id,
+                            ),
+                        )
+                    )
+
+                elif device_rank == 0:
+                    assert response.finish_reason not in (
+                        "error",
+                        "tool_calls",
+                        "function_call",
+                    )
+                    event_sender.send(
+                        ChunkGenerated(
+                            command_id=command_id,
+                            chunk=TokenChunk(
+                                model=shard_metadata.model_card.model_id,
+                                text=response.text,
+                                token_id=response.token,
+                                logprob=response.logprob,
+                                top_logprobs=response.top_logprobs,
+                                finish_reason=response.finish_reason,
+                                stats=response.stats,
+                            ),
+                        )
+                    )
+            case ToolCallResponse():
+                if device_rank == 0:
+                    event_sender.send(
+                        ChunkGenerated(
+                            command_id=command_id,
+                            chunk=ToolCallChunk(
+                                tool_calls=response.tool_calls,
+                                model=shard_metadata.model_card.model_id,
+                            ),
+                        )
+                    )
+
 
 def main(
     bound_instance: BoundInstance,
@@ -101,226 +226,184 @@ def main(
     setup_start_time = time.time()
 
     model: Model | DistributedImageModel | None = None
-    tokenizer = None
+    tokenizer: TokenizerWrapper | None = None
     group = None
+    kv_prefix_cache: KVPrefixCache | None = None
+    batch_handler: BatchedInferenceHandler | None = None
 
     current_status: RunnerStatus = RunnerIdle()
     logger.info("runner created")
     event_sender.send(
         RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
     )
-    with task_receiver as tasks:
-        for task in tasks:
-            event_sender.send(
-                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Running)
-            )
-            event_sender.send(TaskAcknowledged(task_id=task.task_id))
-            match task:
-                case ConnectToGroup() if isinstance(
-                    current_status, (RunnerIdle, RunnerFailed)
-                ):
-                    logger.info("runner connecting")
-                    current_status = RunnerConnecting()
+
+    def process_task(task: Task) -> bool:
+        """
+        Process a single task. Returns True if the runner should continue,
+        False if it should shut down.
+        """
+        nonlocal current_status, model, tokenizer, group, batch_handler
+        event_sender.send(
+            TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Running)
+        )
+        # NOTE: TaskAcknowledged is sent per-case below, AFTER the initial status
+        # update, to avoid a race where the supervisor sees the ack before the
+        # status change and re-dispatches the same lifecycle command.
+        match task:
+            case ConnectToGroup() if isinstance(
+                current_status, (RunnerIdle, RunnerFailed)
+            ):
+                logger.info("runner connecting")
+                current_status = RunnerConnecting()
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
+                    )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
+                group = initialize_mlx(bound_instance)
+
+                logger.info("runner connected")
+                current_status = RunnerConnected()
+
+            # we load the model if it's connected with a group, or idle without a group. we should never tell a model to connect if it doesn't need to
+            case LoadModel() if (
+                isinstance(current_status, RunnerConnected) and group is not None
+            ) or (isinstance(current_status, RunnerIdle) and group is None):
+                current_status = RunnerLoading()
+                logger.info("runner loading")
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
+                    )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
+
+                def on_model_load_timeout() -> None:
                     event_sender.send(
                         RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
+                            runner_id=runner_id,
+                            runner_status=RunnerFailed(
+                                error_message="Model loading timed out"
+                            ),
                         )
                     )
-                    group = initialize_mlx(bound_instance)
+                    time.sleep(0.5)
 
-                    logger.info("runner connected")
-                    current_status = RunnerConnected()
-
-                # we load the model if it's connected with a group, or idle without a group. we should never tell a model to connect if it doesn't need to
-                case LoadModel() if (
-                    isinstance(current_status, RunnerConnected) and group is not None
-                ) or (isinstance(current_status, RunnerIdle) and group is None):
-                    current_status = RunnerLoading()
-                    logger.info("runner loading")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
+                if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
+                    model, tokenizer = load_mlx_items(
+                        bound_instance, group, on_timeout=on_model_load_timeout
                     )
+                    logger.info(f"model has_tool_calling={tokenizer.has_tool_calling}")
 
-                    def on_model_load_timeout() -> None:
-                        event_sender.send(
-                            RunnerStatusUpdated(
-                                runner_id=runner_id,
-                                runner_status=RunnerFailed(
-                                    error_message="Model loading timed out"
-                                ),
-                            )
+                    # Initialize batch handler for text generation models
+                    if BATCH_ENABLED:
+                        # For tensor parallelism, distributed ops are handled inside model layers
+                        # so batch handler should use world_size=1 (no pipelining)
+                        batch_world_size = (
+                            1
+                            if isinstance(shard_metadata, TensorShardMetadata)
+                            else shard_metadata.world_size
                         )
-                        time.sleep(0.5)
-
-                    if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
-                        model, tokenizer = load_mlx_items(
-                            bound_instance, group, on_timeout=on_model_load_timeout
-                        )
-                        logger.info(
-                            f"model has_tool_calling={tokenizer.has_tool_calling}"
-                        )
-                    elif (
-                        ModelTask.TextToImage in shard_metadata.model_card.tasks
-                        or ModelTask.ImageToImage in shard_metadata.model_card.tasks
-                    ):
-                        model = initialize_image_model(bound_instance)
-                    else:
-                        raise ValueError(
-                            f"Unknown model task(s): {shard_metadata.model_card.tasks}"
-                        )
-
-                    current_status = RunnerLoaded()
-                    logger.info("runner loaded")
-                case StartWarmup() if isinstance(current_status, RunnerLoaded):
-                    assert model
-
-                    current_status = RunnerWarmingUp()
-                    logger.info("runner warming up")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
-                    )
-
-                    logger.info(f"warming up inference for instance: {instance}")
-                    if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
-                        assert not isinstance(model, DistributedImageModel)
-                        assert tokenizer
-
-                        toks = warmup_inference(
+                        batch_handler = BatchedInferenceHandler(
                             model=model,
                             tokenizer=tokenizer,
-                            # kv_prefix_cache=kv_prefix_cache,  # supply for warmup-time prefix caching
+                            model_id=shard_metadata.model_card.model_id,
+                            device_rank=device_rank,
+                            world_size=batch_world_size,
+                            max_batch_size=BATCH_MAX_SIZE,
                         )
-                        logger.info(f"warmed up by generating {toks} tokens")
                         logger.info(
-                            f"runner initialized in {time.time() - setup_start_time} seconds"
+                            f"Batch handler initialized (max_batch_size={BATCH_MAX_SIZE}, world_size={batch_world_size})"
                         )
-                    elif (
-                        ModelTask.TextToImage in shard_metadata.model_card.tasks
-                        or ModelTask.ImageToImage in shard_metadata.model_card.tasks
-                    ):
-                        assert isinstance(model, DistributedImageModel)
-                        image = warmup_image_generator(model=model)
-                        if image is not None:
-                            logger.info(f"warmed up by generating {image.size} image")
-                        else:
-                            logger.info("warmup completed (non-primary node)")
+                        kv_prefix_cache = KVPrefixCache(tokenizer)
 
-                    current_status = RunnerReady()
-                    logger.info("runner ready")
-                case ChatCompletion(task_params=task_params, command_id=command_id) if (
-                    isinstance(current_status, RunnerReady)
+                elif (
+                    ModelTask.TextToImage in shard_metadata.model_card.tasks
+                    or ModelTask.ImageToImage in shard_metadata.model_card.tasks
                 ):
-                    logger.info(f"received chat request: {task}")
+                    model = initialize_image_model(bound_instance)
+                else:
+                    raise ValueError(
+                        f"Unknown model task(s): {shard_metadata.model_card.tasks}"
+                    )
+
+                current_status = RunnerLoaded()
+                logger.info("runner loaded")
+            case StartWarmup() if isinstance(current_status, RunnerLoaded):
+                assert model
+
+                current_status = RunnerWarmingUp()
+                logger.info("runner warming up")
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
+                    )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
+
+                logger.info(f"warming up inference for instance: {instance}")
+                if ModelTask.TextGeneration in shard_metadata.model_card.tasks:
+                    assert not isinstance(model, DistributedImageModel)
+                    assert tokenizer
+
+                    toks = warmup_inference(
+                        model=model,
+                        tokenizer=tokenizer,
+                        # kv_prefix_cache=kv_prefix_cache,  # supply for warmup-time prefix caching
+                    )
+                    logger.info(f"warmed up by generating {toks} tokens")
+                    logger.info(
+                        f"runner initialized in {time.time() - setup_start_time} seconds"
+                    )
+                elif (
+                    ModelTask.TextToImage in shard_metadata.model_card.tasks
+                    or ModelTask.ImageToImage in shard_metadata.model_card.tasks
+                ):
+                    assert isinstance(model, DistributedImageModel)
+                    image = warmup_image_generator(model=model)
+                    if image is not None:
+                        logger.info(f"warmed up by generating {image.size} image")
+                    else:
+                        logger.info("warmup completed (non-primary node)")
+
+                current_status = RunnerReady()
+                logger.info("runner ready")
+            case ChatCompletion(task_params=task_params, command_id=command_id) if (
+                isinstance(current_status, (RunnerReady, RunnerRunning))
+            ):
+                logger.info(f"received chat request: {task}")
+                assert model and not isinstance(model, DistributedImageModel)
+                assert tokenizer
+                assert task_params.messages[0].content is not None
+
+                # Check if we should use serial processing for this task
+                if not BATCH_ENABLED:
+                    logger.debug("Serial mode: BATCH_ENABLED is False")
+                    use_serial = True
+                elif batch_handler is None:
+                    logger.debug("Serial mode: batch_handler is None")
+                    use_serial = True
+                else:
+                    use_serial = _should_use_serial_processing(
+                        task, tokenizer, model, shard_metadata.model_card.model_id
+                    )
+
+                if use_serial:
+                    # Serial processing for complex tasks
                     current_status = RunnerRunning()
-                    logger.info("runner running")
+                    logger.info("runner running (serial mode)")
                     event_sender.send(
                         RunnerStatusUpdated(
                             runner_id=runner_id, runner_status=current_status
                         )
                     )
-                    assert model and not isinstance(model, DistributedImageModel)
-                    assert tokenizer
-                    assert task_params.messages[0].content is not None
+                    event_sender.send(TaskAcknowledged(task_id=task.task_id))
 
                     try:
-                        _check_for_debug_prompts(task_params.messages[0].content)
-
-                        # Build prompt once - used for both generation and thinking detection
-                        prompt = apply_chat_template(tokenizer, task_params)
-
-                        # Generate responses using the actual MLX generation
-                        mlx_generator = mlx_generate(
-                            model=model,
-                            tokenizer=tokenizer,
-                            task=task_params,
-                            prompt=prompt,
+                        _process_serial_chat_completion(
+                            task, model, tokenizer, shard_metadata, event_sender
                         )
-
-                        # For other thinking models (GLM, etc.), check if we need to
-                        # prepend the thinking tag that was consumed by the chat template
-                        if detect_thinking_prompt_suffix(prompt, tokenizer):
-                            mlx_generator = parse_thinking_models(
-                                mlx_generator, tokenizer
-                            )
-
-                        # Kimi-K2 has tool call sections - we don't care about them
-                        if "kimi" in shard_metadata.model_card.model_id.lower():
-                            mlx_generator = filter_kimi_tokens(mlx_generator)
-                            patch_kimi_tokenizer(tokenizer)
-
-                        # GLM models need patched parser (upstream has bug with None regex match)
-                        elif "glm" in shard_metadata.model_card.model_id.lower():
-                            patch_glm_tokenizer(tokenizer)
-
-                        # GPT-OSS specific parsing to match other model formats.
-                        elif isinstance(model, GptOssModel):
-                            mlx_generator = parse_gpt_oss(mlx_generator)
-
-                        if tokenizer.has_tool_calling and not isinstance(
-                            model, GptOssModel
-                        ):
-                            assert tokenizer.tool_call_start
-                            assert tokenizer.tool_call_end
-                            assert tokenizer.tool_parser  # pyright: ignore[reportAny]
-                            mlx_generator = parse_tool_calls(
-                                mlx_generator,
-                                tokenizer.tool_call_start,
-                                tokenizer.tool_call_end,
-                                tokenizer.tool_parser,  # pyright: ignore[reportAny]
-                            )
-
-                        for response in mlx_generator:
-                            match response:
-                                case GenerationResponse():
-                                    if (
-                                        device_rank == 0
-                                        and response.finish_reason == "error"
-                                    ):
-                                        event_sender.send(
-                                            ChunkGenerated(
-                                                command_id=command_id,
-                                                chunk=ErrorChunk(
-                                                    error_message=response.text,
-                                                    model=shard_metadata.model_card.model_id,
-                                                ),
-                                            )
-                                        )
-
-                                    elif device_rank == 0:
-                                        assert response.finish_reason not in (
-                                            "error",
-                                            "tool_calls",
-                                            "function_call",
-                                        )
-                                        event_sender.send(
-                                            ChunkGenerated(
-                                                command_id=command_id,
-                                                chunk=TokenChunk(
-                                                    model=shard_metadata.model_card.model_id,
-                                                    text=response.text,
-                                                    token_id=response.token,
-                                                    finish_reason=response.finish_reason,
-                                                    stats=response.stats,
-                                                ),
-                                            )
-                                        )
-                                case ToolCallResponse():
-                                    if device_rank == 0:
-                                        event_sender.send(
-                                            ChunkGenerated(
-                                                command_id=command_id,
-                                                chunk=ToolCallChunk(
-                                                    tool_calls=response.tool_calls,
-                                                    model=shard_metadata.model_card.model_id,
-                                                ),
-                                            )
-                                        )
-
-                    # can we make this more explicit?
                     except Exception as e:
                         if device_rank == 0:
                             event_sender.send(
@@ -337,53 +420,29 @@ def main(
 
                     current_status = RunnerReady()
                     logger.info("runner ready")
-                case ImageGeneration(
-                    task_params=task_params, command_id=command_id
-                ) if isinstance(current_status, RunnerReady):
-                    assert isinstance(model, DistributedImageModel)
-                    logger.info(f"received image generation request: {str(task)[:500]}")
-                    current_status = RunnerRunning()
-                    logger.info("runner running")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
-                    )
-
+                else:
+                    # Batch processing for simple tasks
+                    assert batch_handler is not None
                     try:
-                        # Generate images using the image generation backend
-                        # Track image_index for final images only
-                        image_index = 0
-                        for response in generate_image(model=model, task=task_params):
-                            if (
-                                shard_metadata.device_rank
-                                == shard_metadata.world_size - 1
-                            ):
-                                match response:
-                                    case PartialImageResponse():
-                                        logger.info(
-                                            f"sending partial ImageChunk {response.partial_index}/{response.total_partials}"
-                                        )
-                                        _process_image_response(
-                                            response,
-                                            command_id,
-                                            shard_metadata,
-                                            event_sender,
-                                            image_index,
-                                        )
-                                    case ImageGenerationResponse():
-                                        logger.info("sending final ImageChunk")
-                                        _process_image_response(
-                                            response,
-                                            command_id,
-                                            shard_metadata,
-                                            event_sender,
-                                            image_index,
-                                        )
-                                        image_index += 1
-                    # can we make this more explicit?
+                        _check_for_debug_prompts(task_params.messages[0].content)
+                        batch_handler.add_request(task)
+
+                        # Update status to running if not already
+                        if not isinstance(current_status, RunnerRunning):
+                            current_status = RunnerRunning()
+                            logger.info("runner running (batch mode)")
+                            event_sender.send(
+                                RunnerStatusUpdated(
+                                    runner_id=runner_id, runner_status=current_status
+                                )
+                            )
+                        event_sender.send(TaskAcknowledged(task_id=task.task_id))
+
+                        # Return True to indicate task was added to batch
+                        # (completion will be sent when batch processes)
+                        return True
                     except Exception as e:
-                        if shard_metadata.device_rank == shard_metadata.world_size - 1:
+                        if device_rank == 0:
                             event_sender.send(
                                 ChunkGenerated(
                                     command_id=command_id,
@@ -395,93 +454,235 @@ def main(
                                 )
                             )
                         raise
-
-                    current_status = RunnerReady()
-                    logger.info("runner ready")
-                case ImageEdits(task_params=task_params, command_id=command_id) if (
-                    isinstance(current_status, RunnerReady)
-                ):
-                    assert isinstance(model, DistributedImageModel)
-                    logger.info(f"received image edits request: {str(task)[:500]}")
-                    current_status = RunnerRunning()
-                    logger.info("runner running")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
-                        )
+            case ImageGeneration(task_params=task_params, command_id=command_id) if (
+                isinstance(current_status, RunnerReady)
+            ):
+                assert isinstance(model, DistributedImageModel)
+                logger.info(f"received image generation request: {str(task)[:500]}")
+                current_status = RunnerRunning()
+                logger.info("runner running")
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
                     )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
 
-                    try:
-                        image_index = 0
-                        for response in generate_image(model=model, task=task_params):
-                            if (
-                                shard_metadata.device_rank
-                                == shard_metadata.world_size - 1
-                            ):
-                                match response:
-                                    case PartialImageResponse():
-                                        logger.info(
-                                            f"sending partial ImageChunk {response.partial_index}/{response.total_partials}"
-                                        )
-                                        _process_image_response(
-                                            response,
-                                            command_id,
-                                            shard_metadata,
-                                            event_sender,
-                                            image_index,
-                                        )
-                                    case ImageGenerationResponse():
-                                        logger.info("sending final ImageChunk")
-                                        _process_image_response(
-                                            response,
-                                            command_id,
-                                            shard_metadata,
-                                            event_sender,
-                                            image_index,
-                                        )
-                                        image_index += 1
-                    except Exception as e:
+                try:
+                    # Generate images using the image generation backend
+                    # Track image_index for final images only
+                    image_index = 0
+                    for response in generate_image(model=model, task=task_params):
                         if shard_metadata.device_rank == shard_metadata.world_size - 1:
+                            match response:
+                                case PartialImageResponse():
+                                    logger.info(
+                                        f"sending partial ImageChunk {response.partial_index}/{response.total_partials}"
+                                    )
+                                    _process_image_response(
+                                        response,
+                                        command_id,
+                                        shard_metadata,
+                                        event_sender,
+                                        image_index,
+                                    )
+                                case ImageGenerationResponse():
+                                    logger.info("sending final ImageChunk")
+                                    _process_image_response(
+                                        response,
+                                        command_id,
+                                        shard_metadata,
+                                        event_sender,
+                                        image_index,
+                                    )
+                                    image_index += 1
+                # can we make this more explicit?
+                except Exception as e:
+                    if shard_metadata.device_rank == shard_metadata.world_size - 1:
+                        event_sender.send(
+                            ChunkGenerated(
+                                command_id=command_id,
+                                chunk=ErrorChunk(
+                                    model=shard_metadata.model_card.model_id,
+                                    finish_reason="error",
+                                    error_message=str(e),
+                                ),
+                            )
+                        )
+                    raise
+
+                current_status = RunnerReady()
+                logger.info("runner ready")
+            case ImageEdits(task_params=task_params, command_id=command_id) if (
+                isinstance(current_status, RunnerReady)
+            ):
+                assert isinstance(model, DistributedImageModel)
+                logger.info(f"received image edits request: {str(task)[:500]}")
+                current_status = RunnerRunning()
+                logger.info("runner running")
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
+                    )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
+
+                try:
+                    image_index = 0
+                    for response in generate_image(model=model, task=task_params):
+                        if shard_metadata.device_rank == shard_metadata.world_size - 1:
+                            match response:
+                                case PartialImageResponse():
+                                    logger.info(
+                                        f"sending partial ImageChunk {response.partial_index}/{response.total_partials}"
+                                    )
+                                    _process_image_response(
+                                        response,
+                                        command_id,
+                                        shard_metadata,
+                                        event_sender,
+                                        image_index,
+                                    )
+                                case ImageGenerationResponse():
+                                    logger.info("sending final ImageChunk")
+                                    _process_image_response(
+                                        response,
+                                        command_id,
+                                        shard_metadata,
+                                        event_sender,
+                                        image_index,
+                                    )
+                                    image_index += 1
+                except Exception as e:
+                    if shard_metadata.device_rank == shard_metadata.world_size - 1:
+                        event_sender.send(
+                            ChunkGenerated(
+                                command_id=command_id,
+                                chunk=ErrorChunk(
+                                    model=shard_metadata.model_card.model_id,
+                                    finish_reason="error",
+                                    error_message=str(e),
+                                ),
+                            )
+                        )
+                    raise
+
+                current_status = RunnerReady()
+                logger.info("runner ready")
+            case Shutdown():
+                if batch_handler is not None:
+                    batch_handler.close()
+                    batch_handler = None
+                current_status = RunnerShuttingDown()
+                logger.info("runner shutting down")
+                event_sender.send(
+                    RunnerStatusUpdated(
+                        runner_id=runner_id, runner_status=current_status
+                    )
+                )
+                event_sender.send(TaskAcknowledged(task_id=task.task_id))
+                current_status = RunnerShutdown()
+            case _:
+                raise ValueError(
+                    f"Received {task.__class__.__name__} outside of state machine in {current_status=}"
+                )
+        event_sender.send(
+            TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Complete)
+        )
+        event_sender.send(
+            RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
+        )
+        return not isinstance(current_status, RunnerShutdown)
+
+    # Track tasks that were added to batch (need completion after batch processes)
+    batched_task_ids: list[tuple[Task, bool]] = []  # (task, completed)
+
+    with task_receiver as tasks:
+        while True:
+            # Check if batch handler is active and needs processing
+            if batch_handler is not None and (
+                batch_handler.is_active or batch_handler.has_pending
+            ):
+                # Drain all available tasks before stepping
+                should_break = False
+                while True:
+                    try:
+                        task = tasks.receive_nowait()
+                        if isinstance(task, ChatCompletion) and isinstance(
+                            current_status, (RunnerReady, RunnerRunning)
+                        ):
+                            was_batched = process_task(task)
+                            if was_batched:
+                                batched_task_ids.append((task, False))
+                        else:
+                            should_continue = process_task(task)
+                            if not should_continue:
+                                should_break = True
+                                break
+                    except WouldBlock:
+                        break  # No more tasks available
+                    except EndOfStream:
+                        should_break = True
+                        break
+                if should_break:
+                    break
+
+                # Flush all pending requests before stepping
+                if batch_handler.has_pending:
+                    logger.info(
+                        f"Flushing batch (pending={len(batch_handler.pending)}, active={batch_handler.current_batch_size})"
+                    )
+                    batch_handler.flush()
+
+                # Step generation and emit events
+                if batch_handler.is_active:
+                    event_count = 0
+                    for event in batch_handler.step():
+                        event_sender.send(event)
+                        event_count += 1
+                    if event_count > 0:
+                        logger.debug(f"Emitted {event_count} events from batch")
+
+                # Check for completed batched tasks
+                if not batch_handler.is_active and not batch_handler.has_pending:
+                    # All batched tasks completed
+                    for task, completed in batched_task_ids:
+                        if not completed:
                             event_sender.send(
-                                ChunkGenerated(
-                                    command_id=command_id,
-                                    chunk=ErrorChunk(
-                                        model=shard_metadata.model_card.model_id,
-                                        finish_reason="error",
-                                        error_message=str(e),
-                                    ),
+                                TaskStatusUpdated(
+                                    task_id=task.task_id,
+                                    task_status=TaskStatus.Complete,
                                 )
                             )
-                        raise
+                    batched_task_ids.clear()
 
-                    current_status = RunnerReady()
-                    logger.info("runner ready")
-                case Shutdown():
-                    current_status = RunnerShuttingDown()
-                    logger.info("runner shutting down")
-                    event_sender.send(
-                        RunnerStatusUpdated(
-                            runner_id=runner_id, runner_status=current_status
+                    # Return to ready state
+                    if isinstance(current_status, RunnerRunning):
+                        current_status = RunnerReady()
+                        logger.info("runner ready (batch completed)")
+                        event_sender.send(
+                            RunnerStatusUpdated(
+                                runner_id=runner_id, runner_status=current_status
+                            )
                         )
-                    )
-                    current_status = RunnerShutdown()
-                case _:
-                    raise ValueError(
-                        f"Received {task.__class__.__name__} outside of state machine in {current_status=}"
-                    )
-            event_sender.send(
-                TaskStatusUpdated(task_id=task.task_id, task_status=TaskStatus.Complete)
-            )
-            event_sender.send(
-                RunnerStatusUpdated(runner_id=runner_id, runner_status=current_status)
-            )
-            if isinstance(current_status, RunnerShutdown):
-                del model, tokenizer, group
-                mx.clear_cache()
-                import gc
+            else:
+                # No active batch - use blocking receive
+                try:
+                    task = tasks.receive()
+                    should_continue = process_task(task)
+                    if not should_continue:
+                        break
+                except EndOfStream:
+                    break
 
-                gc.collect()
-                break
+        # Cleanup
+        if batch_handler is not None:
+            batch_handler.close()
+        del model, tokenizer, group
+        mx.clear_cache()
+        import gc
+
+        gc.collect()
 
 
 @cache

src/exo/worker/runner/runner_supervisor.py

@@ -52,6 +52,9 @@ class RunnerSupervisor:
     _tg: TaskGroup | None = field(default=None, init=False)
     status: RunnerStatus = field(default_factory=RunnerIdle, init=False)
     pending: dict[TaskId, anyio.Event] = field(default_factory=dict, init=False)
+    sent: set[TaskId] = field(
+        default_factory=set, init=False
+    )  # Tasks sent to runner (not yet completed)
     completed: set[TaskId] = field(default_factory=set, init=False)
 
     @classmethod
@@ -126,21 +129,39 @@ class RunnerSupervisor:
         assert self._tg
         self._tg.cancel_scope.cancel()
 
-    async def start_task(self, task: Task):
+    async def start_task(self, task: Task, wait_for_ack: bool = True):
+        """
+        Send a task to the runner.
+
+        Args:
+            task: The task to send.
+            wait_for_ack: If True, wait for TaskAcknowledged before returning.
+                          If False, return immediately after sending (for batching).
+        """
         if task.task_id in self.completed:
-            logger.info(
-                f"Skipping invalid task {task} as it has already been completed"
+            logger.debug(
+                f"Skipping task {task.task_id} as it has already been completed"
             )
+            return
+        if task.task_id in self.sent:
+            logger.debug(f"Task {task.task_id} already sent, skipping duplicate")
+            return
+        if task.task_id in self.pending:
+            logger.debug(f"Task {task.task_id} already pending, skipping duplicate")
+            return
         logger.info(f"Starting task {task}")
         event = anyio.Event()
         self.pending[task.task_id] = event
+        self.sent.add(task.task_id)
         try:
             self._task_sender.send(task)
         except ClosedResourceError:
             logger.warning(f"Task {task} dropped, runner closed communication.")
+            self.sent.discard(task.task_id)
             return
-        await event.wait()
-        logger.info(f"Finished task {task}")
+        if wait_for_ack:
+            await event.wait()
+            logger.info(f"Finished task {task}")
 
     async def _forward_events(self):
         with self._ev_recv as events:
@@ -149,7 +170,11 @@ class RunnerSupervisor:
                     if isinstance(event, RunnerStatusUpdated):
                         self.status = event.runner_status
                     if isinstance(event, TaskAcknowledged):
-                        self.pending.pop(event.task_id).set()
+                        # Use pop with default to handle tasks sent with wait_for_ack=False
+                        # that may have already been removed or never added
+                        pending_event = self.pending.pop(event.task_id, None)
+                        if pending_event:
+                            pending_event.set()
                         continue
                     if (
                         isinstance(event, TaskStatusUpdated)
@@ -167,6 +192,7 @@ class RunnerSupervisor:
                             ),
                         )
                         self.completed.add(event.task_id)
+                        self.sent.discard(event.task_id)
                     await self._event_sender.send(event)
             except (ClosedResourceError, BrokenResourceError) as e:
                 await self._check_runner(e)

src/exo/worker/tests/unittests/conftest.py

@@ -20,6 +20,7 @@ class FakeRunnerSupervisor:
     bound_instance: BoundInstance
     status: RunnerStatus
     completed: set[TaskId] = field(default_factory=set)
+    sent: set[TaskId] = field(default_factory=set)
 
 
 class OtherTask(BaseTask):

src/exo/worker/tests/unittests/test_mlx/test_kv_prefix_cache.py

@@ -0,0 +1,545 @@
+# type: ignore
+import time
+from typing import cast
+from unittest.mock import patch
+
+import mlx.core as mx
+import pytest
+from mlx_lm.models.cache import KVCache
+from mlx_lm.sample_utils import make_sampler
+
+from exo.shared.types.api import ChatCompletionMessage
+from exo.shared.types.common import ModelId
+from exo.shared.types.tasks import ChatCompletionTaskParams
+from exo.worker.engines.mlx import Model
+from exo.worker.engines.mlx.cache import (
+    KVPrefixCache,
+    _cache_length,
+    _get_prefix_length,
+    encode_prompt,
+    make_kv_cache,
+)
+from exo.worker.engines.mlx.generator.generate import mlx_generate, prefill
+from exo.worker.engines.mlx.utils_mlx import apply_chat_template
+from exo.worker.tests.unittests.test_mlx.conftest import (
+    DEFAULT_GPT_OSS_CONFIG,
+    DEFAULT_GPT_OSS_MODEL_ID,
+)
+
+
+def _check_model_exists() -> bool:
+    return DEFAULT_GPT_OSS_CONFIG.model_path.exists()
+
+
+class TestGetPrefixLength:
+    def test_identical_arrays(self):
+        a = mx.array([1, 2, 3, 4, 5])
+        b = mx.array([1, 2, 3, 4, 5])
+        assert _get_prefix_length(a, b) == 5
+
+    def test_no_common_prefix(self):
+        a = mx.array([1, 2, 3])
+        b = mx.array([4, 5, 6])
+        assert _get_prefix_length(a, b) == 0
+
+    def test_partial_prefix(self):
+        a = mx.array([1, 2, 3, 4, 5])
+        b = mx.array([1, 2, 3, 7, 8])
+        assert _get_prefix_length(a, b) == 3
+
+    def test_prompt_longer_than_cached(self):
+        a = mx.array([1, 2, 3, 4, 5])
+        b = mx.array([1, 2, 3])
+        assert _get_prefix_length(a, b) == 3
+
+    def test_cached_longer_than_prompt(self):
+        a = mx.array([1, 2, 3])
+        b = mx.array([1, 2, 3, 4, 5])
+        assert _get_prefix_length(a, b) == 3
+
+    def test_single_token_match(self):
+        a = mx.array([1, 2, 3])
+        b = mx.array([1, 5, 6])
+        assert _get_prefix_length(a, b) == 1
+
+    def test_empty_prompt(self):
+        a = mx.array([]).astype(mx.int32)
+        b = mx.array([1, 2, 3])
+        assert _get_prefix_length(a, b) == 0
+
+    def test_empty_cached(self):
+        a = mx.array([1, 2, 3])
+        b = mx.array([]).astype(mx.int32)
+        assert _get_prefix_length(a, b) == 0
+
+    def test_both_empty(self):
+        a = mx.array([]).astype(mx.int32)
+        b = mx.array([]).astype(mx.int32)
+        assert _get_prefix_length(a, b) == 0
+
+
+class TestKVPrefix:
+    @pytest.fixture
+    def mock_tokenizer(self):
+        """Create a minimal mock tokenizer for tests that don't need real tokenization."""
+        from unittest.mock import MagicMock
+
+        tokenizer = MagicMock()
+        tokenizer.encode.return_value = [1, 2, 3]
+        return tokenizer
+
+    def test_starts_empty(self, mock_tokenizer):
+        cache = KVPrefixCache(mock_tokenizer)
+        assert len(cache.prompts) == 0
+        assert len(cache.caches) == 0
+
+    def test_clear_empties_cache(self, mock_tokenizer):
+        cache = KVPrefixCache(mock_tokenizer)
+        cache.prompts.append(mx.array([1, 2, 3]))
+        cache.caches.append([KVCache()])
+        cache.clear()
+        assert len(cache.prompts) == 0
+        assert len(cache.caches) == 0
+
+    def test_clear_on_empty_cache(self, mock_tokenizer):
+        cache = KVPrefixCache(mock_tokenizer)
+        cache.clear()
+        assert len(cache.prompts) == 0
+
+
+def _load_gpt_oss() -> tuple[Model, object]:
+    from mlx_lm.utils import load_model
+
+    from exo.worker.engines.mlx.utils_mlx import load_tokenizer_for_model_id
+
+    model_path = DEFAULT_GPT_OSS_CONFIG.model_path
+    model_id = ModelId(DEFAULT_GPT_OSS_MODEL_ID)
+
+    model, _ = load_model(model_path, lazy=False)
+    tokenizer = load_tokenizer_for_model_id(model_id, model_path)
+    return cast(Model, model), tokenizer
+
+
+@pytest.mark.slow
+@pytest.mark.skipif(
+    not _check_model_exists(),
+    reason=f"GPT-OSS model not found at {DEFAULT_GPT_OSS_CONFIG.model_path}",
+)
+class TestKVPrefixCacheWithModel:
+    @pytest.fixture(scope="class")
+    def model_and_tokenizer(self):
+        model, tokenizer = _load_gpt_oss()
+        return model, tokenizer
+
+    def test_prefill_populates_cache(self, model_and_tokenizer):
+        model, tokenizer = model_and_tokenizer
+
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Hello!!")],
+            max_tokens=1,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        tokens = encode_prompt(tokenizer, prompt)
+        cache = make_kv_cache(model)
+
+        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+
+        # Cache should now hold the prompt tokens
+        assert _cache_length(cache) == len(tokens)
+
+    def test_add_and_get_exact_match(self, model_and_tokenizer):
+        model, tokenizer = model_and_tokenizer
+
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Test exact")],
+            max_tokens=1,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        tokens = encode_prompt(tokenizer, prompt)
+        cache = make_kv_cache(model)
+
+        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache.add_kv_cache(prompt, cache)
+
+        assert len(kv_prefix_cache.prompts) == 1
+        stored_length = _cache_length(kv_prefix_cache.caches[0])
+        assert stored_length > 0
+
+        # Retrieve with same prompt: exact match
+        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
+            model, prompt
+        )
+        assert matched_index == 0
+
+        # Exact match returns only last token
+        assert len(remaining_tokens) == 1
+        assert mx.array_equal(remaining_tokens, tokens[-1:])
+
+    def test_add_and_get_prefix_match(self, model_and_tokenizer):
+        """get_kv_cache with a longer prompt sharing prefix should return partial match."""
+        model, tokenizer = model_and_tokenizer
+
+        short_task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Hi")],
+            max_tokens=1,
+        )
+        short_prompt = apply_chat_template(tokenizer, short_task)
+        short_tokens = encode_prompt(tokenizer, short_prompt)
+        cache = make_kv_cache(model)
+
+        prefill(model, tokenizer, make_sampler(0.0), short_tokens, cache)
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache.add_kv_cache(short_prompt, cache)
+
+        # Query with longer prompt that shares the chat template prefix
+        long_task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[
+                ChatCompletionMessage(role="user", content="Hi there, how are you?")
+            ],
+            max_tokens=1,
+        )
+        long_prompt = apply_chat_template(tokenizer, long_task)
+        long_tokens = encode_prompt(tokenizer, long_prompt)
+
+        # The prompts share a prefix (chat template preamble + "Hi")
+        expected_prefix = _get_prefix_length(long_tokens, short_tokens)
+        assert expected_prefix > 0, (
+            "Prompts should share a prefix from the chat template"
+        )
+
+        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
+            model, long_prompt
+        )
+        assert matched_index == 0
+
+        # remaining_tokens should be the suffix after the shared prefix
+        assert len(remaining_tokens) == len(long_tokens) - expected_prefix
+        assert mx.array_equal(remaining_tokens, long_tokens[expected_prefix:])
+
+    def test_stored_cache_not_mutated_after_get_and_generation(
+        self, model_and_tokenizer
+    ):
+        """Getting a cache and then mutating it (as generation does) must not corrupt stored cache."""
+        model, tokenizer = model_and_tokenizer
+
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Mutation test")],
+            max_tokens=1,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        tokens = encode_prompt(tokenizer, prompt)
+        cache = make_kv_cache(model)
+
+        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache.add_kv_cache(prompt, cache)
+
+        stored_length = _cache_length(kv_prefix_cache.caches[0])
+
+        # Get cache and mutate it (simulating what generation does)
+        result_cache, _, matched_index = kv_prefix_cache.get_kv_cache(model, prompt)
+        assert matched_index == 0
+
+        # Simulate generation: feed many additional tokens through the cache
+        head_dim = result_cache[0].keys.shape[-1]
+        num_heads = result_cache[0].keys.shape[1]
+        extra_keys = mx.random.normal((1, num_heads, 50, head_dim))
+        extra_values = mx.random.normal((1, num_heads, 50, head_dim))
+        for layer_cache in result_cache:
+            layer_cache.update_and_fetch(extra_keys, extra_values)
+        mx.eval([c.keys for c in result_cache])
+
+        # Stored cache must be unchanged
+        assert _cache_length(kv_prefix_cache.caches[0]) == stored_length
+
+    def test_stored_cache_survives_repeated_get_mutate_cycles(
+        self, model_and_tokenizer
+    ):
+        """Multiple get+mutate cycles (like repeated user requests) must not corrupt cache."""
+        model, tokenizer = model_and_tokenizer
+
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Repeat test")],
+            max_tokens=1,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        tokens = encode_prompt(tokenizer, prompt)
+        cache = make_kv_cache(model)
+
+        prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        kv_prefix_cache.add_kv_cache(prompt, cache)
+
+        stored_length = _cache_length(kv_prefix_cache.caches[0])
+
+        for i in range(3):
+            result_cache, _, _ = kv_prefix_cache.get_kv_cache(model, prompt)
+
+            head_dim = result_cache[0].keys.shape[-1]
+            num_heads = result_cache[0].keys.shape[1]
+            extra = mx.random.normal((1, num_heads, 30, head_dim))
+            for layer_cache in result_cache:
+                layer_cache.update_and_fetch(extra, extra)
+            mx.eval([c.keys for c in result_cache])
+
+            assert _cache_length(kv_prefix_cache.caches[0]) == stored_length, (
+                f"Failed on loop {i}"
+            )
+
+    def test_mlx_generate_populates_cache(self, model_and_tokenizer):
+        """mlx_generate should save the cache after generation completes."""
+        model, tokenizer = model_and_tokenizer
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Hello")],
+            max_tokens=5,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        prompt_tokens = encode_prompt(tokenizer, prompt)
+
+        # Consume the entire generator so the cache-saving code after yield runs
+        generated_tokens = 0
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            generated_tokens += 1
+
+        assert len(kv_prefix_cache.prompts) == 1
+        assert len(kv_prefix_cache.caches) == 1
+        # Cache should contain prompt + generated tokens
+        expected_length = len(prompt_tokens) + generated_tokens
+        assert _cache_length(kv_prefix_cache.caches[0]) == expected_length
+
+    def test_mlx_generate_second_call_gets_prefix_hit(self, model_and_tokenizer):
+        """Second mlx_generate call with same prompt should get a prefix hit from stored cache."""
+        model, tokenizer = model_and_tokenizer
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Reuse test")],
+            max_tokens=5,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+        prompt_tokens = encode_prompt(tokenizer, prompt)
+
+        # First generation populates cache
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            pass
+
+        assert len(kv_prefix_cache.prompts) == 1
+
+        # Second call should find a prefix match (the stored cache contains
+        # prompt + generated tokens, which shares the prompt prefix)
+        result_cache, remaining_tokens, matched_index = kv_prefix_cache.get_kv_cache(
+            model, prompt
+        )
+        # The stored cache is longer than the prompt (it includes generated tokens),
+        # so this is a prefix match where our prompt is fully contained
+        assert matched_index == 0
+        # Exact match: remaining_tokens is just the last token
+        assert len(remaining_tokens) == 1
+        assert mx.array_equal(remaining_tokens, prompt_tokens[-1:])
+
+    def test_mlx_generate_long_prompt_updates_cache_in_place(self, model_and_tokenizer):
+        """With a prompt > 1000 tokens, second generation should update the cache entry in-place."""
+        model, tokenizer = model_and_tokenizer
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+
+        # Build a long user message (> 1000 tokens) to exceed _MIN_PREFIX_HIT_TO_UPDATE
+        base_text = "The quick brown fox jumps over the lazy dog. "
+        base_tokens = tokenizer.encode(base_text)
+        repeats = (1200 // len(base_tokens)) + 2
+        long_content = base_text * repeats
+
+        task1 = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content=long_content)],
+            max_tokens=5,
+        )
+        prompt1 = apply_chat_template(tokenizer, task1)
+        prompt1_tokens = encode_prompt(tokenizer, prompt1)
+        assert len(prompt1_tokens) > 1000, (
+            "Prompt must exceed _MIN_PREFIX_HIT_TO_UPDATE"
+        )
+
+        # First generation populates the cache (must prefill all tokens)
+        t0 = time.perf_counter()
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task1,
+            prompt=prompt1,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            pass
+        first_gen_time = time.perf_counter() - t0
+
+        assert len(kv_prefix_cache.prompts) == 1
+        first_cache_length = _cache_length(kv_prefix_cache.caches[0])
+
+        # Second generation: same long prompt + extra content (simulating multi-turn)
+        task2 = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[
+                ChatCompletionMessage(role="user", content=long_content),
+                ChatCompletionMessage(role="assistant", content="Sure, I can help."),
+                ChatCompletionMessage(role="user", content="Tell me more."),
+            ],
+            max_tokens=5,
+        )
+        prompt2 = apply_chat_template(tokenizer, task2)
+        prompt2_tokens = encode_prompt(tokenizer, prompt2)
+
+        # Verify the prompts share a long prefix
+        prefix_len = _get_prefix_length(prompt2_tokens, prompt1_tokens)
+        assert prefix_len > 1000, "Prompts must share > 1000 token prefix"
+
+        # Second generation should reuse the cached prefix (only prefill new tokens)
+        t0 = time.perf_counter()
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task2,
+            prompt=prompt2,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            pass
+        second_gen_time = time.perf_counter() - t0
+
+        # Second generation should be significantly faster due to prefix cache hit - hopefully not flaky
+        assert second_gen_time < first_gen_time * 0.5, (
+            f"Expected prefix cache speedup: "
+            f"first={first_gen_time:.2f}s, second={second_gen_time:.2f}s"
+        )
+
+        # With prefix_hit > 1000, should update in-place (not add a second entry)
+        assert len(kv_prefix_cache.prompts) == 1
+        # Updated cache should be longer (prompt2 + generated > prompt1 + generated)
+        updated_cache_length = _cache_length(kv_prefix_cache.caches[0])
+        assert updated_cache_length > first_cache_length
+
+    def test_mlx_generate_stored_cache_not_mutated(self, model_and_tokenizer):
+        """After mlx_generate saves a cache, a second generation must not corrupt the stored copy."""
+        model, tokenizer = model_and_tokenizer
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+        task = ChatCompletionTaskParams(
+            model=DEFAULT_GPT_OSS_MODEL_ID,
+            messages=[ChatCompletionMessage(role="user", content="Immutable test")],
+            max_tokens=5,
+        )
+        prompt = apply_chat_template(tokenizer, task)
+
+        # First generation populates cache
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            pass
+
+        first_cache_length = _cache_length(kv_prefix_cache.caches[0])
+
+        # Second generation gets the cache and mutates it during generation
+        for _response in mlx_generate(
+            model=model,
+            tokenizer=tokenizer,
+            task=task,
+            prompt=prompt,
+            kv_prefix_cache=kv_prefix_cache,
+        ):
+            pass
+
+        # The first stored cache must not have been mutated by the second generation
+        assert _cache_length(kv_prefix_cache.caches[0]) == first_cache_length
+
+    def test_evicts_lru_entry_under_memory_pressure(self, model_and_tokenizer):
+        """Under memory pressure, adding a new cache entry evicts the least recently used one."""
+        model, tokenizer = model_and_tokenizer
+
+        kv_prefix_cache = KVPrefixCache(tokenizer)
+
+        # Add three cache entries with different prompts
+        prompts = ["First entry", "Second entry", "Third entry"]
+        for i, content in enumerate(prompts):
+            task = ChatCompletionTaskParams(
+                model=DEFAULT_GPT_OSS_MODEL_ID,
+                messages=[ChatCompletionMessage(role="user", content=content)],
+                max_tokens=1,
+            )
+            prompt = apply_chat_template(tokenizer, task)
+            tokens = encode_prompt(tokenizer, prompt)
+            cache = make_kv_cache(model)
+            prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+            kv_prefix_cache.add_kv_cache(prompt, cache)
+            # Stagger _last_used so LRU order is deterministic
+            kv_prefix_cache._last_used[i] = float(i)
+
+        assert len(kv_prefix_cache.prompts) == 3
+
+        # Access the third entry to make it most recently used
+        kv_prefix_cache._last_used[2] = 100.0
+        # Entry 0 (_last_used=0.0) is LRU, entry 1 (_last_used=1.0) is next
+
+        # Simulate memory pressure: active memory exceeds threshold
+        fake_limit = 1000
+        fake_active = int(fake_limit * 0.90)  # Above _MEMORY_THRESHOLD (0.85)
+
+        with (
+            patch(
+                "exo.worker.engines.mlx.cache.mx.metal.get_active_memory",
+                return_value=fake_active,
+            ),
+            patch(
+                "exo.worker.engines.mlx.cache.mx.metal.device_info",
+                return_value={"max_recommended_working_set_size": fake_limit},
+            ),
+        ):
+            # Trigger eviction by adding a new entry
+            task = ChatCompletionTaskParams(
+                model=DEFAULT_GPT_OSS_MODEL_ID,
+                messages=[ChatCompletionMessage(role="user", content="New entry")],
+                max_tokens=1,
+            )
+            prompt = apply_chat_template(tokenizer, task)
+            tokens = encode_prompt(tokenizer, prompt)
+            cache = make_kv_cache(model)
+            prefill(model, tokenizer, make_sampler(0.0), tokens, cache)
+            kv_prefix_cache.add_kv_cache(prompt, cache)
+
+        # LRU entries should have been evicted (entries 0, 1, 2 in order of _last_used)
+        # Since fake_active stays above threshold after each eviction (we don't change it),
+        # all old entries get evicted, leaving only the newly added one
+        assert len(kv_prefix_cache.prompts) == 1
+        # The surviving entry should be the newly added one
+        new_tokens = encode_prompt(tokenizer, prompt)
+        assert _get_prefix_length(kv_prefix_cache.prompts[0], new_tokens) == len(
+            new_tokens
+        )

src/exo/worker/tests/unittests/test_runner/test_event_ordering.py

@@ -118,6 +118,10 @@ def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
     # Returns a prompt without thinking tag so detect_thinking_prompt_suffix returns None.
     monkeypatch.setattr(mlx_runner, "apply_chat_template", make_nothin("test prompt"))
     monkeypatch.setattr(mlx_runner, "detect_thinking_prompt_suffix", make_nothin(False))
+    # Force serial processing mode since batch mode requires a real tokenizer
+    monkeypatch.setattr(mlx_runner, "_should_use_serial_processing", make_nothin(True))
+    # Disable batch handler initialization
+    monkeypatch.setattr(mlx_runner, "BATCH_ENABLED", False)
 
     def fake_generate(*_1: object, **_2: object):
         yield GenerationResponse(token=0, text="hi", finish_reason="stop")
@@ -192,29 +196,30 @@ def test_events_processed_in_correct_order(patch_out_mlx: pytest.MonkeyPatch):
             TaskStatusUpdated(
                 task_id=INITIALIZATION_TASK_ID, task_status=TaskStatus.Running
             ),
-            TaskAcknowledged(task_id=INITIALIZATION_TASK_ID),
+            # Status update comes before ack to prevent race conditions
             RunnerStatusUpdated(
                 runner_id=RUNNER_1_ID, runner_status=RunnerConnecting()
             ),
+            TaskAcknowledged(task_id=INITIALIZATION_TASK_ID),
             TaskStatusUpdated(
                 task_id=INITIALIZATION_TASK_ID, task_status=TaskStatus.Complete
             ),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerConnected()),
             TaskStatusUpdated(task_id=LOAD_TASK_ID, task_status=TaskStatus.Running),
-            TaskAcknowledged(task_id=LOAD_TASK_ID),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerLoading()),
+            TaskAcknowledged(task_id=LOAD_TASK_ID),
             TaskStatusUpdated(task_id=LOAD_TASK_ID, task_status=TaskStatus.Complete),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerLoaded()),
             TaskStatusUpdated(task_id=WARMUP_TASK_ID, task_status=TaskStatus.Running),
-            TaskAcknowledged(task_id=WARMUP_TASK_ID),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerWarmingUp()),
+            TaskAcknowledged(task_id=WARMUP_TASK_ID),
             TaskStatusUpdated(task_id=WARMUP_TASK_ID, task_status=TaskStatus.Complete),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerReady()),
             TaskStatusUpdated(
                 task_id=CHAT_COMPLETION_TASK_ID, task_status=TaskStatus.Running
             ),
-            TaskAcknowledged(task_id=CHAT_COMPLETION_TASK_ID),
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerRunning()),
+            TaskAcknowledged(task_id=CHAT_COMPLETION_TASK_ID),
             expected_chunk,
             TaskStatusUpdated(
                 task_id=CHAT_COMPLETION_TASK_ID, task_status=TaskStatus.Complete
@@ -222,10 +227,10 @@ def test_events_processed_in_correct_order(patch_out_mlx: pytest.MonkeyPatch):
             # CHAT COMPLETION TASK SHOULD COMPLETE BEFORE RUNNER READY
             RunnerStatusUpdated(runner_id=RUNNER_1_ID, runner_status=RunnerReady()),
             TaskStatusUpdated(task_id=SHUTDOWN_TASK_ID, task_status=TaskStatus.Running),
-            TaskAcknowledged(task_id=SHUTDOWN_TASK_ID),
             RunnerStatusUpdated(
                 runner_id=RUNNER_1_ID, runner_status=RunnerShuttingDown()
             ),
+            TaskAcknowledged(task_id=SHUTDOWN_TASK_ID),
             TaskStatusUpdated(
                 task_id=SHUTDOWN_TASK_ID, task_status=TaskStatus.Complete
             ),