Hoist W4A8 activation quantization out of GEMM K-loop#19209
Open
Gasoonjia wants to merge 7 commits intogh/digantdesai/53/basefrom
Open
Hoist W4A8 activation quantization out of GEMM K-loop#19209Gasoonjia wants to merge 7 commits intogh/digantdesai/53/basefrom
Gasoonjia wants to merge 7 commits intogh/digantdesai/53/basefrom
Conversation
… runner Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
…Qwen3.5 MoE runner" Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
…Qwen3.5 MoE runner" Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
…Qwen3.5 MoE runner" Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
…Qwen3.5 MoE runner" Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
…Qwen3.5 MoE runner" Runner now uses llm::Stats with proper timestamps for model load, prefill, decode, and GPU memory (via cudaMemGetInfo). Output matches stats.h print_report format: PyTorchObserver JSON line plus human-readable table. This commit was authored with the assistance of Claude Code. [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19209
Note: Links to docs will display an error until the docs builds have been completed. ❌ 9 New Failures, 2 Cancelled Jobs, 1 Pending, 2 Unrelated FailuresAs of commit d936717 with merge base cb4e5ae ( NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
BROKEN TRUNK - The following job failed but was present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Gasoonjia
force-pushed
the
hoist-activation-quant
branch
2 times, most recently
from
385bf6d to
793be8d
Compare
Add dedicated _quantize_activations_int8_kernel and _silu_quantize_int8_kernel that pre-quantize activations to INT8 with per-row-per-tile FP32 scales before GEMM1 and GEMM2 respectively. The existing _fused_moe_batched_int8_kernel and _fused_moe_silu_batched_int8_kernel are rewritten to consume pre-quantized activations + scales, eliminating ~256 redundant tl.max reductions per program (cdiv(K, BLOCK_K) tiles * BLOCK_M rows) and halving activation HBM bandwidth in the K-loop (bf16 -> int8). BLOCK_SIZE_K is fixed at PREQUANT_BLOCK_K (= 128) so per-tile activation scales align with the GEMM K-loop. Correctness: 7/7 microbenchmark configs pass with rel diff <1.5% vs BF16 ref. End-to-end (Qwen3.5 MoE 1600 prefill + 512 decode, --cuda_graph, A100): prefill 5727 -> 6171 tok/s (+7.7%), decode 92.6 -> 99.0 tok/s (+6.9%).
Gasoonjia
force-pushed
the
hoist-activation-quant
branch
from
793be8d to
d936717
Compare
Base automatically changed from
gh/digantdesai/53/head
to
gh/digantdesai/53/base
April 30, 2026 15:05
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Context
The original K-loop did
tl.max(tl.abs(a))+ INT8 cast on every tile (16 tiles × 16 rows = 256 reductions per program). Hoisting eliminates this redundant work and halves activation HBM bandwidth in the GEMM (bf16 → int8).Improvement
Pre-quantize activations to INT8 once into a dedicated buffer (with per-row-per-tile FP32 scales) before the W4A8 batched MoE GEMMs, instead of re-quantizing inside the K-loop on every tile.
Perf (1600-token prefill)
gh/digantdesai/53/head)Correctness
7/7 microbenchmark configs (incl. qwen3.5-like M=128, K=2048, gs=128) pass with relative diff <1.5% vs BF16 reference — within INT8 quantization noise.