Repository navigation
Conversation
…tivations contiguously The prefill grouped GEMM read each e2m1 code through a table gather and loaded the activations at stride two for the even and odd k of a byte. On an RTX 3090 Ti at Qwen3.8-Flash-Next's shapes (8192 tokens, 512 experts, top-10) that held it at 8.6 TFLOPS; placing the codes in fp16 bits and splitting one contiguous activation tile in registers reaches ~50, with bit-identical output.
…ode with the prefill GEMM
…th compiles at a ragged K
…t placement and load activations contiguously # Conflicts: # tests/kernels/test_e4m3_compat.py
This was referenced Oct 2, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stacked on #77. This merges upstream FlashML-org#578 at head
e428b06. It changes the NVFP4 MoE prefill GEMM (the Triton grouped kernel, which is what prefill runs on sm_120 here):EVEN_Kpath.Upstream reports bit-identical outputs.
Conflict:
tests/kernels/test_e4m3_compat.py. Both sides add cases to_emit_all, and both are kept.Measured (RTX 5080, Qwen3.8-Flash-Next-NVFP4, production config at
2224380), interleaved base/FlashML-org#578/base/FlashML-org#578, uncached prompts, time to first token:* Only the second round is shown at 4K, because each arm's first 4K request includes the server's first-prefill kernel compile.
Bit-identity: greedy fingerprints on 3 fixed prompts are identical (2 rounds), and 20 cold nonce prompts × 3 give 0/60 mismatches.
Tests:
test_nvfp4_backends.py,test_e4m3_compat.py,test_fused_copy.py: 38 passed.Deployed.
🤖 Generated with Claude Code
https://claude.ai/code/session_01Bu6LgoxLR4wETqb7RPR2vt