[maithilijoshi20]

Fixes #35252

Summary

• Keep the int4_w4a16 tuner config key aligned with the runtime down-projection dimension.
• Add regression coverage for the emitted config filename.

Testing

• Focused MoE config test: 3 passed.
• Ruff, isort, Black, and git diff --check passed.
• CUDA/Triton smoke check passed on an NVIDIA RTX 3050 in WSL.

The config-only test uses a stub for sgl_kernel because the available binary wheel targets a different GPU architecture; no kernel execution is covered by this change.

---

CI States

Latest PR Test (Base): ❌ Run #32810867033
Latest PR Test (Extra): ❌ Run #32810866824
Latest PR Test (AMD ROCm 7.2): ❌ Run #32810867026

[maithilijoshi20]

@JeremiahM37 This PR removes the extra int4-specific halving from both the tuner lookup and generated config filename, so the tuner now emits the runtime key (for example, E=256,N=256). Would you be able to try your reproduction and confirm it resolves the issue on your setup?

[JeremiahM37]

Confirmed, this fixes it — thanks for picking it up.

I couldn't re-run the full tune+serve (my ROCm image is gone), so I verified the filename logic directly against the model from the report (E=256, shard_intermediate_size=512):

main     -> E=256,N=128,...,dtype=int4_w4a16.json
this PR  -> E=256,N=256,...,dtype=int4_w4a16.json

which is exactly the key the server was looking for.

Good catch on the second copy in common_utils.py — I'd only cited the one in tuning_fused_moe_triton.py, and fixing that alone would have left the written filename wrong.

[maithilijoshi20]

@BBuf Could you help review or route this small fused-MoE tuner config fix? The issue reporter validated the filename correction.

[BBuf]

@BBuf Could you help review or route this small fused-MoE tuner config fix? The issue reporter validated the filename correction.

Thanks for your fix.