Repository navigation
perf(moe): preserve squared-ReLU fused MoE kernel behind a dedicated flag (#268) - #280
Merged
Merged
Conversation
…flag (#268) nemotron-h's MoE experts are fc1 -> relu² -> fc2 (squared ReLU, not SwiGLU). A fused decode path for them was built and validated byte-identical, but it is performance-NEUTRAL on nemotron-h-30b (54.9 -> 54.7 tok/s): nemotron's decode is dominated by its 23 Mamba2 + 6 attention layers, so the MoE expert GEMV is a small, already-efficient slice. The finding is that nemotron-h's 0.60x decode gap is the hybrid SSM/attention path, not the MoE expert kernel. Rather than discard the kernel, keep it in the tree behind its own flag MLXCEL_FUSED_MOE_RELU2 (separate from the default MLXCEL_FUSED_MOE, which stays on gather_qmm for nemotron — no regression). The fc1+relu² kernel produces act_g[K, Dff]; fc2 reuses moe_down_kernel. Revive it for a future squared-ReLU MoE model whose decode is MoE-dominated rather than Mamba-hybrid. Verified: default flag leaves nemotron on gather_qmm (byte-identical to off); MLXCEL_FUSED_MOE_RELU2 path byte-identical to the reference. fmt + clippy clean.
inureyes
added
type:performance
This was referenced Jun 14, 2026
inureyes
added a commit
that referenced
this pull request
Jun 14, 2026
…#283) * docs(moe): document MLXCEL_FUSED_MOE flags, per-model gains, and M5 follow-up (#268) Add a usage/flags section to the fused decode-MoE design doc: the MLXCEL_FUSED_MOE / MLXCEL_FUSED_MOE_SGY / MLXCEL_FUSED_MOE_RELU2 env vars, the measured per-model decode gains on M1 Ultra (gemma4 +13%, qwen3.5 +8.7%, dots +4.7%, qwen3-30b +3.5%, nemotron ~0%), the f16-jitter-class parity caveat, and the list of covered models. Mark roadmap steps 3-4 done (#278/#279/#280/#281) and add step 5: validate on M5, then decide on flipping the flag default-on. * docs(moe): reference the M5 validation issue (#282)
This was referenced Jun 14, 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
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.
Summary
Step 3b of #268 investigated wiring nemotron-h's MoE into the fused decode kernel. nemotron-h's experts are fc1 → relu² → fc2 (squared ReLU, not SwiGLU), so a dedicated fc1+relu² kernel was built (fc2 reuses the existing
moe_down_kernel).Finding: it's correct but performance-neutral on nemotron-h. Byte-identical output, but decode is unchanged/slightly lower (54.9 → 54.7 tok/s). nemotron-h's decode is dominated by its 23 Mamba2 + 6 attention layers, so the MoE expert GEMV is a small, already-efficient slice. nemotron-h's 0.60× gap is the hybrid SSM/attention path, not the MoE expert kernel.
What this PR does
Rather than discard a correct kernel, keep it in the tree behind its own flag
MLXCEL_FUSED_MOE_RELU2, separate from the defaultMLXCEL_FUSED_MOE(which stays ongather_qmmfor nemotron — no regression). This preserves the kernel as referenceable, compilable, testable code for a future squared-ReLU MoE model whose decode is MoE-dominated rather than Mamba-hybrid.Validation
MLXCEL_FUSED_MOE: nemotron-h output byte-identical to fused-off (stays on gather_qmm, no behavior change).MLXCEL_FUSED_MOE_RELU2: output byte-identical to the reference (kernel correct).cargo fmt --check+cargo clippyclean.Refs #268.