[WIP] Persistent + grouped tile MoE batched GEMM #19219
Draft
Gasoonjia wants to merge 1 commit intohoist-activation-quantfrom
Draft
[WIP] Persistent + grouped tile MoE batched GEMM #19219Gasoonjia wants to merge 1 commit intohoist-activation-quantfrom
Gasoonjia wants to merge 1 commit intohoist-activation-quantfrom
Conversation
…-quant
Rewrites the four batched MoE GEMM kernels (BF16/INT8 GEMM1 + GEMM2) from
one-CTA-per-tile to a persistent grid: launch min(NUM_SMS, num_tiles) programs,
each loops over its share of (expert_block, n_block) tiles via
`tl.range(start_pid, num_tiles, NUM_SMS)`.
Tiles are visited in column-major-within-group order (Triton tutorial style)
so consecutive M-blocks of the same expert reuse B[expert, n_block, *] via L2.
Since moe_align_block_size already sorts blocks by expert, this gives clean
weight reuse across the group.
Adds NUM_SMS (constexpr) and GROUP_SIZE_M (autotuned over {8, 16, 32}) to
all four batched kernels and their wrappers. NUM_SMS is queried once per
device via _num_sms() with lru_cache.
Benchmark on A100 80GB, prefill=1642, decode=512, --cuda_graph
(persistent_on_hoist vs hoist baseline):
hoist baseline: 6171 prefill (5941-6313), 99.0 decode tok/s
persistent + grouped: 6127 prefill (6037-6315), 84.5 decode tok/s
Best-case prefill is essentially tied with the hoist baseline at this
prefill length on A100 (108 SMs). The hoist optimization had already
shortened the K-loop by removing per-tile activation quantization, so the
persistent kernel's L2 reuse gain doesn't show on top of it for M=1642.
The optimization should matter more for short prefills (M <= 512) and on
GPUs with fewer SMs / bigger L2 (e.g. RTX 4090 with 128 SMs and 72MB L2)
where wave quantization is more severe.
Behavior preserved: kernels are mathematically identical to the hoist
baseline; only the tile traversal order and grid launch shape changed.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19219
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New Failures, 3 Cancelled Jobs, 3 Pending, 1 Unrelated FailureAs of commit a67881e 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:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Gasoonjia
force-pushed
the
persistent-on-hoist
branch
from
015476d to
a67881e
Compare
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
1 participant
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.
No description provided.