Add Triton INT4 dense kernels with dequant prefill path for Qwen3.5 MoE#19188
Open
digantdesai wants to merge 5 commits intogh/digantdesai/51/basefrom
Open
Add Triton INT4 dense kernels with dequant prefill path for Qwen3.5 MoE#19188digantdesai wants to merge 5 commits intogh/digantdesai/51/basefrom
digantdesai wants to merge 5 commits intogh/digantdesai/51/basefrom
Conversation
Add three new Triton kernels for dense W4A16 linear projections that replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights (same format as MoE experts): - int4_matmul: fused dequant+tl.dot GEMM for medium M (prefill crossover) - int4_matvec: bandwidth-optimized vec-mat for M=1 decode - dequant_w4_to_bf16: weight dequant for large-M prefill via Inductor mm W4DequantLinear wraps these with dual decode/prefill dispatch: - Decode (M=1): int4_matvec (73 tok/s, ~35% slower than tinygemm) - Prefill (M>1): dequant+F.linear via Inductor (3400 tok/s at 3K tokens, +67% over tinygemm baseline) Single 18GB weight blob (no duplication). Decode perf regression is a known trade-off for uniform weight format — to be revisited with a CUDA C++ matvec kernel. Also adds INT8 dynamic-activation MoE tests and comprehensive correctness tests (48 tests, all passing at rtol=0.01). Co-authored-by: Claude <noreply@anthropic.com> [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19188
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New Failures, 4 Cancelled Jobs, 2 Unrelated FailuresAs of commit 21c2f4b with merge base cb4e5ae ( NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
BROKEN TRUNK - The following jobs failed but were 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
This was referenced Apr 28, 2026
This PR needs a
|
…r Qwen3.5 MoE" Add three new Triton kernels for dense W4A16 linear projections that replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights (same format as MoE experts): - int4_matmul: fused dequant+tl.dot GEMM for medium M (prefill crossover) - int4_matvec: bandwidth-optimized vec-mat for M=1 decode - dequant_w4_to_bf16: weight dequant for large-M prefill via Inductor mm W4DequantLinear wraps these with dual decode/prefill dispatch: - Decode (M=1): int4_matvec (73 tok/s, ~35% slower than tinygemm) - Prefill (M>1): dequant+F.linear via Inductor (3400 tok/s at 3K tokens, +67% over tinygemm baseline) Single 18GB weight blob (no duplication). Decode perf regression is a known trade-off for uniform weight format — to be revisited with a CUDA C++ matvec kernel. Also adds INT8 dynamic-activation MoE tests and comprehensive correctness tests (48 tests, all passing at rtol=0.01). Co-authored-by: Claude <noreplyanthropic.com> [ghstack-poisoned]
…r Qwen3.5 MoE" Add three new Triton kernels for dense W4A16 linear projections that replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights (same format as MoE experts): - int4_matmul: fused dequant+tl.dot GEMM for medium M (prefill crossover) - int4_matvec: bandwidth-optimized vec-mat for M=1 decode - dequant_w4_to_bf16: weight dequant for large-M prefill via Inductor mm W4DequantLinear wraps these with dual decode/prefill dispatch: - Decode (M=1): int4_matvec (73 tok/s, ~35% slower than tinygemm) - Prefill (M>1): dequant+F.linear via Inductor (3400 tok/s at 3K tokens, +67% over tinygemm baseline) Single 18GB weight blob (no duplication). Decode perf regression is a known trade-off for uniform weight format — to be revisited with a CUDA C++ matvec kernel. Also adds INT8 dynamic-activation MoE tests and comprehensive correctness tests (48 tests, all passing at rtol=0.01). Co-authored-by: Claude <noreplyanthropic.com> [ghstack-poisoned]
…r Qwen3.5 MoE" Add three new Triton kernels for dense W4A16 linear projections that replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights (same format as MoE experts): - int4_matmul: fused dequant+tl.dot GEMM for medium M (prefill crossover) - int4_matvec: bandwidth-optimized vec-mat for M=1 decode - dequant_w4_to_bf16: weight dequant for large-M prefill via Inductor mm W4DequantLinear wraps these with dual decode/prefill dispatch: - Decode (M=1): int4_matvec (73 tok/s, ~35% slower than tinygemm) - Prefill (M>1): dequant+F.linear via Inductor (3400 tok/s at 3K tokens, +67% over tinygemm baseline) Single 18GB weight blob (no duplication). Decode perf regression is a known trade-off for uniform weight format — to be revisited with a CUDA C++ matvec kernel. Also adds INT8 dynamic-activation MoE tests and comprehensive correctness tests (48 tests, all passing at rtol=0.01). Co-authored-by: Claude <noreplyanthropic.com> [ghstack-poisoned]
Gasoonjia
approved these changes
Apr 29, 2026
Contributor
Gasoonjia
left a comment
Gasoonjia
left a comment
There was a problem hiding this comment.
Please update ci to cover the new changes
…r Qwen3.5 MoE" Add three new Triton kernels for dense W4A16 linear projections that replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights (same format as MoE experts): - int4_matmul: fused dequant+tl.dot GEMM for medium M (prefill crossover) - int4_matvec: bandwidth-optimized vec-mat for M=1 decode - dequant_w4_to_bf16: weight dequant for large-M prefill via Inductor mm W4DequantLinear wraps these with dual decode/prefill dispatch: - Decode (M=1): int4_matvec (73 tok/s, ~35% slower than tinygemm) - Prefill (M>1): dequant+F.linear via Inductor (3400 tok/s at 3K tokens, +67% over tinygemm baseline) Single 18GB weight blob (no duplication). Decode perf regression is a known trade-off for uniform weight format — to be revisited with a CUDA C++ matvec kernel. Also adds INT8 dynamic-activation MoE tests and comprehensive correctness tests (48 tests, all passing at rtol=0.01). Co-authored-by: Claude <noreplyanthropic.com> [ghstack-poisoned]
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.
Stack from ghstack (oldest at bottom):
Add three new Triton kernels for dense W4A16 linear projections that
replace tinygemm's tiled INT4 format with simple [N, K//2] packed weights
(same format as MoE experts):
W4DequantLinear wraps these with dual decode/prefill dispatch:
+67% over tinygemm baseline)
Single 18GB weight blob (no duplication). Decode perf regression is a
known trade-off for uniform weight format — to be revisited with a
CUDA C++ matvec kernel.
Also adds INT8 dynamic-activation MoE tests and comprehensive correctness
tests (48 tests, all passing at rtol=0.01).
Co-authored-by: Claude noreply@anthropic.com