[None][perf] NVFP4 linear-SF AllReduce epilogue - #17452
Draft
brb-nv wants to merge 3 commits into
Draft
Conversation
…ar SF layout The fused epilogue already emits NVFP4, but only with scale factors in the SWIZZLED 128x4 layout the CUTLASS FP4 GEMMs want. The trtllm-gen MoE wants LINEAR row-major and asserts on a swizzled Fp4QuantizedTensor, so no producer could hand it a pre-quantized activation. Less was missing than it looked. The AR fusion kernel already takes a layout field, and cvt_quant_get_sf_out_offset already has a LINEAR branch which, with the arguments that kernel passes, reduces to rowIdx * numColVecs + kIdx. Only the host side hardcoded swizzled: the layout was never set and scale_out was always sized with computeSwizzledLayoutSFSize. Exposed as a new fusion op rather than a layout flag. A flag would have had to cross four op schemas -- allreduce, allreduce_pg, tunable_allreduce and their fakes -- and every Python caller; an op is additive, changes no existing caller, and the UB and MNNVL paths reject it for free because they allow-list the ops they support. Two things fell out. selectStrategyLookUpTable dereferenced mapFusionOpToIndex.find(op) without checking it, so any op missing from that table was undefined behavior under AUTO; the new op is now in the table and the lookup falls back to NCCL on a miss. And TRTLLMGenFusedMoE.forward_impl opened with an assert on x.dtype despite typing x as Union[Tensor, Fp4QuantizedTensor], which has no dtype -- the accepted type had never been exercised on that backend. Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>
…ce epilogue Between the attention o_proj AllReduce and the first routed-expert GEMM sits a serial window on the main stream, every sparse layer, and one of the kernels in it is the standalone NVFP4 quantize of the MoE input. The AllReduce that opens the window already computes the bf16 RMSNorm result that quantize reads, so it can emit the quantized activation from the same pass. Uses the OUT_ variant because the router gate and the shared expert both still need the bf16 norm. The fold is bitwise-identical: the epilogue quantizes from the same bf16 values it writes to norm_out, through the same conversion the standalone kernel uses. It stands down when the experts are not NVFP4, when NVFP4_AWQ needs a per-channel pre-quant scale the epilogue cannot apply, and when the MoE would pad the activation past the hidden size. Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>
Asserts what makes the fold safe to fold: that fp4_quantize(norm_out, scale, 16, False, False) reproduces the epilogue's own fp4 and scale-factor outputs with atol=rtol=0, so the quantized activation is the same one the standalone kernel would have produced from the norm the epilogue writes. One trap is worth recording. This module is registered with cloudpickle.register_pickle_by_value, and cloudpickle then pulls in any sys.modules entry under a referenced module whose name components all appear in the worker's own co_names. Reading torch.ops directly in a worker body drags in sys.modules["torch.ops"], whose type subclasses ModuleType instead of being it, so it misses cloudpickle's module reducer and fails to pickle in the parent -- surfacing as "cannot pickle '_Ops' object" with no rank ever starting. Every worker here reads torch.ops from a nested helper, which keeps the name off the outer code object. Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com>
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.
@coderabbitai summary
Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
If PR introduces API changes, an appropriate PR label is added - either
api-compatibleorapi-breaking. Forapi-breaking, includeBREAKINGin the PR title.Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
To see a list of available CI bot commands, please comment
/bot help.