mirror of
https://github.com/mudler/LocalAI.git
synced 2026-07-03 04:46:54 -04:00
feat(paged): add moe mmq launch trace patch
Assisted-by: Codex:gpt-5
This commit is contained in:
Ettore Di Giacinto
@@ -87,7 +87,7 @@ orthogonal to the paged allocator.
|
||||
|
||||
---
|
||||
|
||||
## 3. Patch series (0001-0055)
|
||||
## 3. Patch series (0001-0057)
|
||||
|
||||
Source-only patches, with intentional numbering gaps (e.g. 0005, 0027). The
|
||||
decode-serving graph-reuse levers are 0040-0041. "Bit-exact" = greedy md5 /
|
||||
@@ -214,6 +214,7 @@ These are the dominant decode levers on the Qwen3.6 hybrid models. All bit-exact
|
||||
| 0054 | **Disable backend sampling for MTP drafts** - forces server MTP draft generation through the target-side sampler acceptance path instead of letting the draft backend sample independently. This was required for the Phase 14 rollback/prefix safety gate. | yes for canonical non-MTP gates; Phase 14 MTP normalized greedy-prefix gate passed |
|
||||
| 0055 | **Trace speculative batch shapes** - adds default-off `LLAMA_SPEC_SHAPE_TRACE=1` server logs around `server_slot::handle_last_sampled_token()`, reporting normal decode rows and MTP verification `K + 1` rows (`draft`, `outputs`, `spec_i_first`, `spec_i_last`). This is instrumentation only for Phase 18 shape-entropy measurement before any scheduler experiment. | yes (env unset is silent; DGX gates after patch: MoE `8cb0ce23`, dense `5951a5b4`, `MUL_MAT_ID` `806/806`) |
|
||||
| 0056 | **Trace MoE MMQ batch shapes** - adds default-off `LLAMA_MOE_MMQ_SHAPE_TRACE=<n>` logs from the grouped-MMQ host selector, reporting routed assignment count, estimated active experts, density, selected `mmq_x`, `mmq_y`, and stream-k. This is evidence-only instrumentation for sizing structural grouped-MMQ work after Phase 28 rejected launch-bounds/row-tile knobs. | yes (env unset and trace-enabled gates both green: MoE `8cb0ce23`, dense `5951a5b4`, `MUL_MAT_ID` `806/806`; trace cap verified with 4 lines) |
|
||||
| 0057 | **Trace MoE MMQ launch shapes** - extends `LLAMA_MOE_MMQ_SHAPE_TRACE=<n>` with bounded `[LLAMA_MOE_MMQ_LAUNCH]` lines from `launch_mul_mat_q`, recording actual `ntiles_dst`, `stream_k_blocks`, tile efficiency, `fixup`, `ntx/nty/ntzw`, and compiled `mmq_x/mmq_y`. This is evidence-only instrumentation to distinguish real stream-k/fixup overhead from small-M kernel-shape cost. | yes (default-off, trace-enabled, and post-serving gates green: MoE `8cb0ce23`, dense `5951a5b4`, `MUL_MAT_ID` `806/806`; Phase 31 n128 trace showed decode and prefill `fixup=0`, `stream_k_blocks == ntiles_dst`) |
|
||||
|
||||
> **Dropped: patch 0026 (hybrid per-head bf16 SSM state, `ssm_bf16_tau`).** Once
|
||||
> the decode fusions (0028 recurrent-state gather-fusion + 0029 block-table cache)
|
||||
@@ -647,3 +648,12 @@ added test-first (`test-cuda-mmq-shape-trace`), compiled under CUDA on DGX, and
|
||||
kept inference stable with the trace disabled and enabled:
|
||||
MoE `8cb0ce23`, dense `5951a5b4`, `MUL_MAT_ID 806/806`. Example trace line:
|
||||
`[LLAMA_MOE_MMQ_SHAPE] type=40 moe=1 ncols_dst=104 nchannels_x=256 ncols_max=13 n_active_est=104 density=1 mmq_x_max=128 mmq_x_lim=64 mmq_x_best=16 mmq_y=128 stream_k=1`.
|
||||
|
||||
Phase 31 extended that trace as patch `0057`
|
||||
(`/home/mudler/bench/phase31_mmq_launch_trace/20260701_064424`) with
|
||||
`[LLAMA_MOE_MMQ_LAUNCH]` lines from `launch_mul_mat_q`. Default-off,
|
||||
trace-enabled, and post-serving gates stayed stable: MoE `8cb0ce23`, dense
|
||||
`5951a5b4`, `MUL_MAT_ID 806/806`. The n128 serving trace showed decode-like
|
||||
`4800/4800` and prefill-like `4920/4920` launch lines with `fixup=0` and
|
||||
`stream_k_blocks == ntiles_dst`, rejecting a no-fixup/no-stream-k shortcut for
|
||||
this workload.
|
||||
|
||||
@@ -1861,3 +1861,49 @@ Decision:
|
||||
separate A/B.
|
||||
- Every traced call used stream-k, so a replacement kernel must account for the
|
||||
current stream-k/fixup behavior rather than only conventional tiling.
|
||||
|
||||
## Phase 31 Live MoE MMQ Launch Shape Distribution
|
||||
|
||||
Phase 31 added patch `0057`, a default-off launch trace paired with the Phase 29
|
||||
selector trace. It records the actual launch policy after `ntiles_dst`,
|
||||
`tiles_efficiency_percent`, `stream_k_blocks`, and `fixup_needed` are known.
|
||||
|
||||
Artifact:
|
||||
|
||||
- `/home/mudler/bench/phase31_mmq_launch_trace/20260701_064424`
|
||||
|
||||
Run:
|
||||
|
||||
- Fork commit: `/home/mudler/_git/llama.cpp` `c78e537b5`
|
||||
- DGX mirror commit: `dgx:~/llama-phase6-source` `8b75905e9`
|
||||
- Env: `LLAMA_KV_PAGED=1 LLAMA_MOE_FORCE_GRAPHS=1 LLAMA_MOE_MMQ_SHAPE_TRACE=4096`
|
||||
- Workload: h2h `n=128`, `PTOK=128`, `GEN=64`
|
||||
- Throughput while tracing: `decode_agg_tps=691.0`, `agg_tps=337.0`,
|
||||
`prefill_tps=1500.4`, `TTFT mean=7671.0 ms`
|
||||
|
||||
Launch summary:
|
||||
|
||||
| bucket | launch lines | `fixup=1` | `stream_k_blocks == ntiles_dst` | tile efficiency | `ncols_max` range |
|
||||
|--------|--------------|-----------|----------------------------------|-----------------|-------------------|
|
||||
| decode-like (`ncols_max <= 128`) | 4800 | 0 | 4800 | 96-99 | 12-128 |
|
||||
| prefill-like (`ncols_max > 128`) | 4920 | 0 | 4920 | 99-100 | 129-510 |
|
||||
|
||||
Gates:
|
||||
|
||||
| check | status | actual |
|
||||
|-------|--------|--------|
|
||||
| default-off MoE md5 | ok | `8cb0ce23777bf55f92f63d0292c756b0` |
|
||||
| default-off dense md5 | ok | `5951a5b4d624ce891e22ab5fca9bc439` |
|
||||
| trace-enabled MoE md5 | ok | `8cb0ce23777bf55f92f63d0292c756b0` |
|
||||
| trace-enabled dense md5 | ok | `5951a5b4d624ce891e22ab5fca9bc439` |
|
||||
| post-serving MoE md5 | ok | `8cb0ce23777bf55f92f63d0292c756b0` |
|
||||
| post-serving dense md5 | ok | `5951a5b4d624ce891e22ab5fca9bc439` |
|
||||
| `MUL_MAT_ID` | ok | `806/806` in all three gate runs |
|
||||
|
||||
Decision:
|
||||
|
||||
- Do not pursue a no-fixup/no-stream-k shortcut for n128 serving: the measured
|
||||
launch path already uses `stream_k_blocks == ntiles_dst` and never runs fixup.
|
||||
- The remaining grouped-MMQ work is structural small-M kernel work, not launch
|
||||
overhead. A follow-up should target the decode-like `mmq_x <= 64`, low-density
|
||||
kernel shape directly and keep the prefill `mmq_x=128` path separate.
|
||||
|
||||
@@ -64,7 +64,7 @@ A lever compiled into the binary is **NOT** isolated by a runtime flag alone. It
|
||||
- **Always update the fork FIRST, in this exact order:** (1) commit the change on the `localai-paged` branch and **push it**, then (2) regenerate the LocalAI series (`backend/cpp/llama-cpp-localai-paged/patches/paged/`) from the fork via `git format-patch` (one patch per fork commit, source-only, never touching a `*.md`/dev-doc), so the series stays a **1:1, drift-free mirror** of the branch. No hand-export.
|
||||
- **NEVER edit the LocalAI `patches/paged/*.patch` files directly**, and **NEVER add a patch to the series with no corresponding fork-branch commit.** They are generated output, not source.
|
||||
- The fork branch is also **where the build and the per-path bit-exact md5 gate actually run**, so it is the **only** place a change is truly validated. A patch that lives only in the LocalAI series has never been built or gated.
|
||||
- **Mirror invariant (verify by tree hash):** applying the full on-disk series on the pin must reproduce the fork branch tree byte-for-byte. The series has **intentional gaps** (missing 0005, 0026, 0027, 0032, 0036-0039, 0045), so the patch count is not the max number; what must hold is the tree-hash equality, not the count. Current verified state: fork HEAD `20a99518a` is mirrored by worktree patch `0056-feat-cuda-trace-moe-mmq-batch-shapes.patch`; applying all `47` patch files on `0ed235ea2c17a19fc8238668653946721ed136fd` produces tree `8a7779726a81689a14f10a64523f2cc380d4801f`, exactly matching the fork.
|
||||
- **Mirror invariant (verify by tree hash):** applying the full on-disk series on the pin must reproduce the fork branch tree byte-for-byte. The series has **intentional gaps** (missing 0005, 0026, 0027, 0032, 0036-0039, 0045), so the patch count is not the max number; what must hold is the tree-hash equality, not the count. Current verified state: fork HEAD `c78e537b5` is mirrored by worktree patch `0057-feat-cuda-trace-moe-mmq-launch-shapes.patch`; applying all `48` patch files on `0ed235ea2c17a19fc8238668653946721ed136fd` produces tree `4eae628e4ba6f2defa14a19d19f7e4abef9a2647`, exactly matching the fork.
|
||||
|
||||
### 2.6 Bench hygiene gates
|
||||
- **NEVER set `LLAMA_MAX_BATCH_TOKENS` in benches** (the harness explicitly logs "NO LLAMA_MAX_BATCH_TOKENS").
|
||||
@@ -322,11 +322,11 @@ Use this harness for future current-stack GB10 snapshots. Do not reuse
|
||||
`~/bench/combined_definitive.sh` unless it is first ported away from stale
|
||||
`~/llama-paged-dev` paths and old lock assumptions.
|
||||
|
||||
Phase 22 re-verified the patch-series mirror invariant after patch `0055`:
|
||||
Phase 31 re-verified the patch-series mirror invariant after patch `0057`:
|
||||
applying every LocalAI `patches/paged/0*.patch` with strict `git apply` on top of
|
||||
Makefile pin `0ed235ea2c17a19fc8238668653946721ed136fd` produced tree
|
||||
`5bdbf8ea3d750fe6fa1f85175fd6357d36222edb`, exactly matching fork branch
|
||||
`localai-paged` HEAD `20a99518a feat(cuda): trace moe mmq batch shapes`.
|
||||
`4eae628e4ba6f2defa14a19d19f7e4abef9a2647`, exactly matching fork branch
|
||||
`localai-paged` HEAD `c78e537b5 feat(cuda): trace moe mmq launch shapes`.
|
||||
|
||||
Phase 24 extended `paged-current-serving-snapshot.sh` to write the snapshot
|
||||
hardware report. DGX dry run passed at
|
||||
@@ -405,6 +405,20 @@ grouped-MMQ calls split into 1200 decode-like calls (`ncols_max <= 128`) and
|
||||
gates stayed green: MoE `8cb0ce23777bf55f92f63d0292c756b0`, dense
|
||||
`5951a5b4d624ce891e22ab5fca9bc439`, `MUL_MAT_ID` `806/806`.
|
||||
|
||||
Phase 31 added patch `0057` for default-off grouped-MMQ launch tracing.
|
||||
Artifact: `/home/mudler/bench/phase31_mmq_launch_trace/20260701_064424`.
|
||||
Fork commit: `c78e537b5 feat(cuda): trace moe mmq launch shapes`; DGX mirror
|
||||
commit: `8b75905e9`. The trace adds `[LLAMA_MOE_MMQ_LAUNCH]` lines under
|
||||
`LLAMA_MOE_MMQ_SHAPE_TRACE=<n>`, recording `ntiles_dst`, `stream_k_blocks`,
|
||||
tile efficiency, `fixup`, `ntx/nty/ntzw`, and compiled `mmq_x/mmq_y`. Default
|
||||
off, trace-enabled, and post-serving gates stayed green: MoE
|
||||
`8cb0ce23777bf55f92f63d0292c756b0`, dense
|
||||
`5951a5b4d624ce891e22ab5fca9bc439`, `MUL_MAT_ID` `806/806`. The n128 serving
|
||||
trace showed decode-like `4800/4800` and prefill-like `4920/4920` launch lines
|
||||
with `fixup=0` and `stream_k_blocks == ntiles_dst`. Do not pursue a
|
||||
no-fixup/no-stream-k shortcut for this workload; the remaining grouped-MMQ work
|
||||
is structural small-M kernel work.
|
||||
|
||||
---
|
||||
|
||||
## 5. METHODOLOGY LESSONS (so you do not repeat the mistakes)
|
||||
@@ -454,15 +468,15 @@ Only pursue if (a)+(b) are not options and someone explicitly wants the residual
|
||||
## 7. KEY FILE / ARTIFACT INDEX
|
||||
|
||||
### Fork (canonical source of truth)
|
||||
- Local canonical fork: `/home/mudler/_git/llama.cpp`, branch **`localai-paged`**, HEAD `20a99518a39acbb4474fa9c97121fc7b9f07c1ef` ("trace moe mmq batch shapes", patch `0056`).
|
||||
- DGX current clean mirror/build tree: `dgx:~/llama-phase6-source`, HEAD `826c97a05` with the Phase 29 shape-trace patch applied and committed; Phase 20/26/27 artifacts still record their historical source hashes.
|
||||
- Local canonical fork: `/home/mudler/_git/llama.cpp`, branch **`localai-paged`**, HEAD `c78e537b56e3446f8aa645c6700aacf263639bd8` ("trace moe mmq launch shapes", patch `0057`).
|
||||
- DGX current clean mirror/build tree: `dgx:~/llama-phase6-source`, HEAD `8b75905e9` with the Phase 31 launch-trace patch applied and committed; Phase 20/26/27 artifacts still record their historical source hashes.
|
||||
- Historical DGX dev tree: `dgx:~/llama-paged-dev`, branch **`paged`**, HEAD `a7d439e8ce6990eb09721223c975da4e49d8d136` ("GDN CONFIG C (M8) - bf16 Kc/Qc"). It is an old experimental tree and must not be treated as canonical.
|
||||
|
||||
### LocalAI worktree
|
||||
- Path: `/home/mudler/_git/LocalAI/.claude/worktrees/feat+paged-attention`, branch `worktree-feat+paged-attention` (currently 246 ahead, 31 behind `origin/master`; recompute before reporting).
|
||||
- Backend dir: `backend/cpp/llama-cpp-localai-paged/` (`Makefile` thin wrapper, `package.sh`, `run.sh`, `README.md` ~44 KB canonical, `docs/`, `patches/paged/`).
|
||||
- `docs/`: `VLLM_PARITY_FINAL.md` (authoritative record), `VLLM_PARITY_LEVER_MAP.md` (working brainstorm, profile-validated section), `DECODE_SERVING_SCOPE.md`, `PREFILL_GEMM_SCOPE.md`, `PREFILL_GEMM_RESULTS.md`, `TENSORCORE_GDN_SCOPE.md`, `TENSORCORE_GDN_BUILD_PLAN.md`, `ACCELERATOR_PORTING_SCOPE.md`, `UPSTREAM_LAYER2_SCOPE.md`, `LOCALAI_LLAMACPP_BACKEND_PLAN.md`, `PAGED_BITEXACT_NOTE.md`, `PATCH_MAINTENANCE.md`, `final_benchmark.csv`, `paged-burst-bench.cpp`, `paged-reclaim-unit.cpp`, 3 PNGs, and this `PARITY_HANDOFF.md`.
|
||||
- `patches/paged/`: **46** `.patch` files spanning 0001-0055 with intentional gaps (missing 0005, 0026 [dropped ssm_bf16_tau], 0027, 0032, 0036-0039, 0045). Core paged-KV 0001-0012; decode-first scheduler 0013/0016; serving graph reuse 0040/0041; prefill fusions 0042/0044; SSM/GDN decode 0018-0022/0028; MoE NVFP4 quant 0023/0025/0043; FP4-MMA/Marlin scaffolds 0033/0034/0035 (default-off); GDN tensor-core prefill 0031 -> 0046 (geometry gate) -> 0047 (f32-only M5, default-on under paged KV); W4A16 packed metadata/shape/padding is 0048-0050; MoE safety tests are 0051-0053; MTP backend-sampling safety is 0054; speculative shape trace is 0055.
|
||||
- `patches/paged/`: **48** `.patch` files spanning 0001-0057 with intentional gaps (missing 0005, 0026 [dropped ssm_bf16_tau], 0027, 0032, 0036-0039, 0045). Core paged-KV 0001-0012; decode-first scheduler 0013/0016; serving graph reuse 0040/0041; prefill fusions 0042/0044; SSM/GDN decode 0018-0022/0028; MoE NVFP4 quant 0023/0025/0043; FP4-MMA/Marlin scaffolds 0033/0034/0035 (default-off); GDN tensor-core prefill 0031 -> 0046 (geometry gate) -> 0047 (f32-only M5, default-on under paged KV); W4A16 packed metadata/shape/padding is 0048-0050; MoE safety tests are 0051-0053; MTP backend-sampling safety is 0054; speculative shape trace is 0055; MoE MMQ selector/launch traces are 0056-0057.
|
||||
|
||||
### Bench artifacts (DGX)
|
||||
- `~/bench/COMBINED_DEFINITIVE.txt` (+ `.log`, `.done`, `combined_definitive.sh`, `combined_definitive.out`) - historical same-session both-engine run.
|
||||
@@ -475,6 +489,7 @@ Only pursue if (a)+(b) are not options and someone explicitly wants the residual
|
||||
- `~/bench/phase28_mmq_occupancy/20260701_040450` - NVFP4 MMQ occupancy build-knob A/B; `MINBLOCKS=2` gate-safe but serving-regressed, `MMQ_Y=64` compile-rejected.
|
||||
- `~/bench/phase29_mmq_shape_trace/20260701_042428` - default-off MoE MMQ shape trace patch `0056`; CUDA build plus default/trace md5 gates green.
|
||||
- `~/bench/phase30_mmq_shape_serving/20260701_043300` - live n128 serving MMQ shape distribution from patch `0056`; post-run md5/op gates green.
|
||||
- `~/bench/phase31_mmq_launch_trace/20260701_064424` - default-off MoE MMQ launch trace patch `0057`; default/trace/post-serving md5 gates green; n128 launch trace rejects stream-k/fixup shortcut (`fixup=0`, `stream_k_blocks == ntiles_dst`).
|
||||
- Per-engine logs `~/bench/COMBINED_{paged,vllm}_{MOE,DENSE}_server.log`; `~/bench/BENCHMARK_PROGRESS.md`.
|
||||
- Graph-node-traced high-N profiles: `~/highN_prof2/*.nsys-rep` (paged npl=256), `~/highN_vllm/*.nsys-rep` (vLLM), 2026-06-30.
|
||||
- A/B dirs: `~/bench/marlin_gate/`, `~/bench/gdn_p1_ab/`.
|
||||
@@ -487,8 +502,8 @@ Only pursue if (a)+(b) are not options and someone explicitly wants the residual
|
||||
|
||||
### Discrepancies to flag / resolve (carried verbatim from the gather, including UNVERIFIED labels)
|
||||
1. **Pin prose reconciled in this worktree.** Makefile line 52 `LLAMA_VERSION?=0ed235ea2c17a19fc8238668653946721ed136fd` is authoritative and matches the local fork merge-base. Hard rule: the paged pin must equal the stock `llama-cpp` pin (shared `grpc-server.cpp`); a bump to `c299a92c` once broke the grpc-server link despite being bit-exact and was reverted. Trust the Makefile when building.
|
||||
2. **Current fork/mirror are clean and verified.** Local fork HEAD is `20a99518a`, DGX clean mirror HEAD is `826c97a05`, and Phase 29 re-proved the LocalAI patch series tree equals the fork tree (`8a7779726a81689a14f10a64523f2cc380d4801f`). The old `llama-paged-dev` tree is historical only.
|
||||
3. **Worktree patch series is tracked through 0056.** The only current untracked path in this worktree is `.claude/`.
|
||||
2. **Current fork/mirror are clean and verified.** Local fork HEAD is `c78e537b5`, DGX clean mirror HEAD is `8b75905e9`, and Phase 31 should be treated as the current patch-series tip. The old `llama-paged-dev` tree is historical only.
|
||||
3. **Worktree patch series is tracked through 0057.** The only expected unrelated untracked path in this worktree is `.claude/`.
|
||||
4. **`sm_121a` is not in the worktree build files** - it lives only in the DGX experimental build scripts (`gdn_cc.sh`, `gdn_bv_build.sh`, `paged-build.sh`); mainline uses arch `121`. **UNVERIFIED** whether the shipped CI Dockerfile build path injects `121a` for the FP4-MMA kernels (`Dockerfile.llama-cpp-localai-paged` does not hardcode a CUDA arch).
|
||||
5. **The `0921716...` paged-MoE md5 open item.** `COMBINED_DEFINITIVE.txt` records `PAGED_GATE_MD5=0921716cd0582b5d15af8c362b811d00` for MoE, but a full doc/patch/`git log -S` grep of the worktree found **no** occurrence of `0921716...` in any committed source; the committed canonical paged-MoE gate is `8cb0ce23`. Treat this as **unreconciled**: the documented, KL-validated paged-MoE gate remains `8cb0ce23`, and any paged-MoE divergence (including `0921716`) must be KL-validated against the f16 reference before being accepted as benign, never on assertion alone. The `0921716` value is **UNVERIFIED** as a sanctioned gate; do not adopt it as canonical without re-running the KL gate. The **dense** run is symmetric: `COMBINED_DEFINITIVE.txt` records `PAGED_GATE_MD5=ecfe924dee6c5622c149f419ff2a6481` for dense, which likewise differs from the canonical dense gate `5951a5b4`. Both CDEF `PAGED_GATE_MD5` values come from the `combined_definitive.sh` harness's own gate command, NOT the canonical bit-exact gate command in section 3.3, which is why they diverge from the committed `8cb0ce23` / `5951a5b4`; neither is a sanctioned gate and both must be KL-validated before being treated as benign.
|
||||
|
||||
|
||||
@@ -57,18 +57,18 @@ everywhere without ever touching the stock `llama-cpp` source tree.
|
||||
|
||||
## Latest mirror check
|
||||
|
||||
Phase 29 re-verified the mirror invariant after adding patch `0056`:
|
||||
Phase 31 re-verified the mirror invariant after adding patch `0057`:
|
||||
|
||||
```text
|
||||
base=0ed235ea2c17a19fc8238668653946721ed136fd
|
||||
applied_tree=8a7779726a81689a14f10a64523f2cc380d4801f
|
||||
fork_tree=8a7779726a81689a14f10a64523f2cc380d4801f
|
||||
applied_tree=4eae628e4ba6f2defa14a19d19f7e4abef9a2647
|
||||
fork_tree=4eae628e4ba6f2defa14a19d19f7e4abef9a2647
|
||||
```
|
||||
|
||||
The check used a fresh worktree at `LLAMA_VERSION`, applied every
|
||||
`patches/paged/0*.patch` with strict `git apply`, staged the result, and compared
|
||||
`git write-tree` to canonical fork branch `localai-paged` at
|
||||
`20a99518a feat(cuda): trace moe mmq batch shapes`.
|
||||
`c78e537b5 feat(cuda): trace moe mmq launch shapes`.
|
||||
|
||||
## Status
|
||||
|
||||
|
||||
@@ -802,6 +802,29 @@ small-M decode tiles (`ncols_max` 26-111, density 1-4) separately from prefill.
|
||||
The current stream-k/fixup path is part of the measured shape and cannot be
|
||||
ignored by a replacement kernel.
|
||||
|
||||
### Phase 31 live serving MMQ launch distribution
|
||||
|
||||
Phase 31 added patch `0057`, extending `LLAMA_MOE_MMQ_SHAPE_TRACE=<n>` with
|
||||
`[LLAMA_MOE_MMQ_LAUNCH]` lines emitted from `launch_mul_mat_q` after the actual
|
||||
stream-k launch policy is known. Artifact:
|
||||
`/home/mudler/bench/phase31_mmq_launch_trace/20260701_064424`.
|
||||
|
||||
The default-off, trace-enabled, and post-serving gates all stayed bit-exact:
|
||||
MoE `8cb0ce23777bf55f92f63d0292c756b0`, dense
|
||||
`5951a5b4d624ce891e22ab5fca9bc439`, and `MUL_MAT_ID` `806/806`.
|
||||
|
||||
Live n128 serving with `LLAMA_MOE_MMQ_SHAPE_TRACE=4096` produced:
|
||||
|
||||
| bucket | launch lines | `fixup=1` | `stream_k_blocks == ntiles_dst` | tile efficiency |
|
||||
|--------|--------------|-----------|----------------------------------|-----------------|
|
||||
| decode-like (`ncols_max <= 128`) | 4800 | 0 | 4800 | 96-99 |
|
||||
| prefill-like (`ncols_max > 128`) | 4920 | 0 | 4920 | 99-100 |
|
||||
|
||||
Lever implication: a no-fixup/no-stream-k shortcut is rejected for the measured
|
||||
n128 serving workload. The launch code is already choosing conventional
|
||||
stream-k tiling with no fixup; the remaining gap is the small-M grouped-MMQ
|
||||
kernel shape itself, not launch/fixup overhead.
|
||||
|
||||
Relevant files (all absolute): `/home/mudler/_git/LocalAI/.claude/worktrees/feat+paged-attention/backend/cpp/llama-cpp-localai-paged/docs/{DECODE_SERVING_SCOPE.md,PREFILL_GEMM_SCOPE.md,PREFILL_GEMM_RESULTS.md,TENSORCORE_GDN_SCOPE.md,final_benchmark.csv}`, `.../README.md`, `.../patches/paged/0034-feat-paged-native-NVFP4-W4A4-FP4-MMA-large-M-prefill.patch` (P1/P2), `.../patches/paged/0042-feat-paged-fused-residual-add-RMS-norm-weight-multip.patch` (P7), `.../patches/paged/0031` (P4), `0025` (D1), `0018/0022` (D4/D5), `0009/0010` (D3/D6/D7); graph source `/home/mudler/_git/LocalAI/backend/cpp/llama-cpp-paged-dev/src/{models/qwen35moe.cpp,models/delta-net-base.cpp,llama-graph.cpp}`.
|
||||
|
||||
### Phase 10 GDN C32 slab update
|
||||
|
||||
@@ -0,0 +1,223 @@
|
||||
From c78e537b56e3446f8aa645c6700aacf263639bd8 Mon Sep 17 00:00:00 2001
|
||||
From: Ettore Di Giacinto <mudler@localai.io>
|
||||
Date: Wed, 1 Jul 2026 04:49:55 +0000
|
||||
Subject: [PATCH] feat(cuda): trace moe mmq launch shapes
|
||||
|
||||
Assisted-by: Codex:gpt-5
|
||||
---
|
||||
ggml/src/ggml-cuda/mmq-shape-trace.h | 65 ++++++++++++++++++++++++++++
|
||||
ggml/src/ggml-cuda/mmq.cuh | 47 +++++++++++++-------
|
||||
tests/test-cuda-mmq-shape-trace.cpp | 35 +++++++++++++++
|
||||
3 files changed, 132 insertions(+), 15 deletions(-)
|
||||
|
||||
diff --git a/ggml/src/ggml-cuda/mmq-shape-trace.h b/ggml/src/ggml-cuda/mmq-shape-trace.h
|
||||
index 9d41b7c80..98bc21f7f 100644
|
||||
--- a/ggml/src/ggml-cuda/mmq-shape-trace.h
|
||||
+++ b/ggml/src/ggml-cuda/mmq-shape-trace.h
|
||||
@@ -19,6 +19,24 @@ struct ggml_cuda_mmq_shape {
|
||||
bool use_stream_k;
|
||||
};
|
||||
|
||||
+struct ggml_cuda_mmq_launch_shape {
|
||||
+ int type;
|
||||
+ bool is_moe;
|
||||
+ int64_t ncols_dst;
|
||||
+ int64_t ncols_max;
|
||||
+ int mmq_x;
|
||||
+ int mmq_y;
|
||||
+ int ntx;
|
||||
+ int nty;
|
||||
+ int ntzw;
|
||||
+ int ntiles_dst;
|
||||
+ int nsm;
|
||||
+ int tiles_nwaves;
|
||||
+ int tiles_efficiency_percent;
|
||||
+ int stream_k_blocks;
|
||||
+ bool fixup_needed;
|
||||
+};
|
||||
+
|
||||
static inline ggml_cuda_mmq_shape ggml_cuda_mmq_shape_make(
|
||||
const int type, const bool is_moe, const int64_t ncols_dst, const int64_t nchannels_x,
|
||||
const int64_t ncols_max, const int mmq_x_max, const int mmq_x_lim, const int mmq_x_best,
|
||||
@@ -46,6 +64,30 @@ static inline ggml_cuda_mmq_shape ggml_cuda_mmq_shape_make(
|
||||
};
|
||||
}
|
||||
|
||||
+static inline ggml_cuda_mmq_launch_shape ggml_cuda_mmq_launch_shape_make(
|
||||
+ const int type, const bool is_moe, const int64_t ncols_dst, const int64_t ncols_max,
|
||||
+ const int mmq_x, const int mmq_y, const int ntx, const int nty, const int ntzw,
|
||||
+ const int ntiles_dst, const int nsm, const int tiles_nwaves, const int tiles_efficiency_percent,
|
||||
+ const int stream_k_blocks, const bool fixup_needed) {
|
||||
+ return {
|
||||
+ type,
|
||||
+ is_moe,
|
||||
+ ncols_dst,
|
||||
+ ncols_max,
|
||||
+ mmq_x,
|
||||
+ mmq_y,
|
||||
+ ntx,
|
||||
+ nty,
|
||||
+ ntzw,
|
||||
+ ntiles_dst,
|
||||
+ nsm,
|
||||
+ tiles_nwaves,
|
||||
+ tiles_efficiency_percent,
|
||||
+ stream_k_blocks,
|
||||
+ fixup_needed,
|
||||
+ };
|
||||
+}
|
||||
+
|
||||
static inline int ggml_cuda_mmq_shape_format(char * buf, const size_t size, const ggml_cuda_mmq_shape & shape) {
|
||||
return std::snprintf(buf, size,
|
||||
"type=%d moe=%d ncols_dst=%lld nchannels_x=%lld ncols_max=%lld "
|
||||
@@ -64,3 +106,26 @@ static inline int ggml_cuda_mmq_shape_format(char * buf, const size_t size, cons
|
||||
shape.mmq_y,
|
||||
shape.use_stream_k ? 1 : 0);
|
||||
}
|
||||
+
|
||||
+static inline int ggml_cuda_mmq_launch_shape_format(
|
||||
+ char * buf, const size_t size, const ggml_cuda_mmq_launch_shape & shape) {
|
||||
+ return std::snprintf(buf, size,
|
||||
+ "type=%d moe=%d ncols_dst=%lld ncols_max=%lld mmq_x=%d mmq_y=%d "
|
||||
+ "ntx=%d nty=%d ntzw=%d ntiles_dst=%d nsm=%d tiles_nwaves=%d "
|
||||
+ "tiles_efficiency=%d stream_k_blocks=%d fixup=%d",
|
||||
+ shape.type,
|
||||
+ shape.is_moe ? 1 : 0,
|
||||
+ (long long) shape.ncols_dst,
|
||||
+ (long long) shape.ncols_max,
|
||||
+ shape.mmq_x,
|
||||
+ shape.mmq_y,
|
||||
+ shape.ntx,
|
||||
+ shape.nty,
|
||||
+ shape.ntzw,
|
||||
+ shape.ntiles_dst,
|
||||
+ shape.nsm,
|
||||
+ shape.tiles_nwaves,
|
||||
+ shape.tiles_efficiency_percent,
|
||||
+ shape.stream_k_blocks,
|
||||
+ shape.fixup_needed ? 1 : 0);
|
||||
+}
|
||||
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
|
||||
index 6bc943738..34002edf7 100644
|
||||
--- a/ggml/src/ggml-cuda/mmq.cuh
|
||||
+++ b/ggml/src/ggml-cuda/mmq.cuh
|
||||
@@ -3989,6 +3989,24 @@ static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int
|
||||
return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int));
|
||||
}
|
||||
|
||||
+static inline int ggml_cuda_moe_mmq_shape_trace_limit() {
|
||||
+ static const int limit = []() -> int {
|
||||
+ const char * s = getenv("LLAMA_MOE_MMQ_SHAPE_TRACE");
|
||||
+ if (s == nullptr || strcmp(s, "0") == 0) {
|
||||
+ return 0;
|
||||
+ }
|
||||
+ const int parsed = atoi(s);
|
||||
+ return parsed > 0 ? parsed : 256;
|
||||
+ }();
|
||||
+ return limit;
|
||||
+}
|
||||
+
|
||||
+static inline bool ggml_cuda_moe_mmq_trace_take(std::atomic<int> & counter) {
|
||||
+ const int trace_limit = ggml_cuda_moe_mmq_shape_trace_limit();
|
||||
+ const int trace_index = trace_limit > 0 ? counter.fetch_add(1, std::memory_order_relaxed) : trace_limit;
|
||||
+ return trace_index >= 0 && trace_index < trace_limit;
|
||||
+}
|
||||
+
|
||||
template <ggml_type type, int mmq_x>
|
||||
static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
|
||||
const int id = ggml_cuda_get_device();
|
||||
@@ -4054,6 +4072,19 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
|
||||
|
||||
const bool fixup_needed = ntiles_dst % block_nums_stream_k.x != 0;
|
||||
|
||||
+ if (args.expert_bounds != nullptr) {
|
||||
+ static std::atomic<int> trace_count{0};
|
||||
+ if (ggml_cuda_moe_mmq_trace_take(trace_count)) {
|
||||
+ char buf[256];
|
||||
+ const ggml_cuda_mmq_launch_shape shape = ggml_cuda_mmq_launch_shape_make(
|
||||
+ (int) type, true, args.ncols_dst, args.ncols_max, mmq_x, mmq_y,
|
||||
+ ntx, nty, ntzw, ntiles_dst, nsm, tiles_nwaves, tiles_efficiency_percent,
|
||||
+ block_nums_stream_k.x, fixup_needed);
|
||||
+ ggml_cuda_mmq_launch_shape_format(buf, sizeof(buf), shape);
|
||||
+ fprintf(stderr, "[LLAMA_MOE_MMQ_LAUNCH] %s\n", buf);
|
||||
+ }
|
||||
+ }
|
||||
+
|
||||
ggml_cuda_pool & pool = ctx.pool(id);
|
||||
ggml_cuda_pool_alloc<float> tmp_fixup(pool);
|
||||
if (fixup_needed) {
|
||||
@@ -4167,18 +4198,6 @@ static inline int ggml_cuda_fp4_dense_mmq_x_cap() {
|
||||
return c;
|
||||
}
|
||||
|
||||
-static inline int ggml_cuda_moe_mmq_shape_trace_limit() {
|
||||
- static const int limit = []() -> int {
|
||||
- const char * s = getenv("LLAMA_MOE_MMQ_SHAPE_TRACE");
|
||||
- if (s == nullptr || strcmp(s, "0") == 0) {
|
||||
- return 0;
|
||||
- }
|
||||
- const int parsed = atoi(s);
|
||||
- return parsed > 0 ? parsed : 256;
|
||||
- }();
|
||||
- return limit;
|
||||
-}
|
||||
-
|
||||
template <ggml_type type>
|
||||
void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
|
||||
const int id = ggml_cuda_get_device();
|
||||
@@ -4267,9 +4286,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
|
||||
|
||||
if (args.expert_bounds != nullptr) {
|
||||
static std::atomic<int> trace_count{0};
|
||||
- const int trace_limit = ggml_cuda_moe_mmq_shape_trace_limit();
|
||||
- const int trace_index = trace_limit > 0 ? trace_count.fetch_add(1, std::memory_order_relaxed) : trace_limit;
|
||||
- if (trace_index >= 0 && trace_index < trace_limit) {
|
||||
+ if (ggml_cuda_moe_mmq_trace_take(trace_count)) {
|
||||
char buf[256];
|
||||
const ggml_cuda_mmq_shape shape = ggml_cuda_mmq_shape_make(
|
||||
(int) type, true, args.ncols_dst, args.nchannels_x, args.ncols_max,
|
||||
diff --git a/tests/test-cuda-mmq-shape-trace.cpp b/tests/test-cuda-mmq-shape-trace.cpp
|
||||
index 8620169c0..86ee15e02 100644
|
||||
--- a/tests/test-cuda-mmq-shape-trace.cpp
|
||||
+++ b/tests/test-cuda-mmq-shape-trace.cpp
|
||||
@@ -38,5 +38,40 @@ int main() {
|
||||
require(std::strstr(buf, "mmq_x_best=64") != nullptr, "trace includes selected tile");
|
||||
require(std::strstr(buf, "stream_k=1") != nullptr, "trace includes stream-k flag");
|
||||
|
||||
+ const ggml_cuda_mmq_launch_shape launch = ggml_cuda_mmq_launch_shape_make(
|
||||
+ /* type */ 39,
|
||||
+ /* is_moe */ true,
|
||||
+ /* ncols_dst */ 1024,
|
||||
+ /* ncols_max */ 128,
|
||||
+ /* mmq_x */ 64,
|
||||
+ /* mmq_y */ 128,
|
||||
+ /* ntx */ 2,
|
||||
+ /* nty */ 4,
|
||||
+ /* ntzw */ 3,
|
||||
+ /* ntiles_dst */ 24,
|
||||
+ /* nsm */ 16,
|
||||
+ /* tiles_nwaves */ 2,
|
||||
+ /* tiles_efficiency_percent */ 75,
|
||||
+ /* stream_k_blocks */ 16,
|
||||
+ /* fixup_needed */ true);
|
||||
+
|
||||
+ const int launch_n = ggml_cuda_mmq_launch_shape_format(buf, sizeof(buf), launch);
|
||||
+
|
||||
+ require(launch_n > 0, "launch format returns byte count");
|
||||
+ require(std::strstr(buf, "moe=1") != nullptr, "launch trace includes moe flag");
|
||||
+ require(std::strstr(buf, "ncols_dst=1024") != nullptr, "launch trace includes routed assignment count");
|
||||
+ require(std::strstr(buf, "ncols_max=128") != nullptr, "launch trace includes max column count");
|
||||
+ require(std::strstr(buf, "mmq_x=64") != nullptr, "launch trace includes compiled x tile");
|
||||
+ require(std::strstr(buf, "mmq_y=128") != nullptr, "launch trace includes compiled y tile");
|
||||
+ require(std::strstr(buf, "ntx=2") != nullptr, "launch trace includes x tile count");
|
||||
+ require(std::strstr(buf, "nty=4") != nullptr, "launch trace includes y tile count");
|
||||
+ require(std::strstr(buf, "ntzw=3") != nullptr, "launch trace includes batch tile count");
|
||||
+ require(std::strstr(buf, "ntiles_dst=24") != nullptr, "launch trace includes total tile count");
|
||||
+ require(std::strstr(buf, "nsm=16") != nullptr, "launch trace includes SM count");
|
||||
+ require(std::strstr(buf, "tiles_nwaves=2") != nullptr, "launch trace includes wave count");
|
||||
+ require(std::strstr(buf, "tiles_efficiency=75") != nullptr, "launch trace includes stream-k efficiency");
|
||||
+ require(std::strstr(buf, "stream_k_blocks=16") != nullptr, "launch trace includes actual stream-k block count");
|
||||
+ require(std::strstr(buf, "fixup=1") != nullptr, "launch trace includes fixup flag");
|
||||
+
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user