feat: INT32 atomic MoE kernel for deterministic accumulation

[Mortis-Huang]

Summary

• Replace global_atomic_pk_add_bf16 with global_atomic_add (INT32) in MoE kernel to achieve bit-identical results across runs
• Add INT32 post-processing in fused_moe.py (scale factor 32768, interleaved layout decode)
• Force kernel selection to the modified non-PS non-tkw1 variant to avoid unmodified kernel paths

How to Enable

export AITER_MOE_INT32_ATOMIC=1

Files Changed

File	Description
aiter/fused_moe.py	INT32 post-processing + forced kernel selection logic
hsa/gfx942/fmoe/silu/fmoe_bf16_pertokenFp8_g1u1_vs_silu_1tg_32x192.co	INT32 atomic kernel binary (scale=32768)

Test Results

Consistency (bit-identical)

• 10/10 runs identical across all 131,072 output elements
• Verified with verify_int32_consistency.py

GSM8K Accuracy (parallel=64, 1319 questions)

Metric	BF16 Baseline	INT32	Diff
Accuracy	94.6%	94.5%	-0.1%
Invalid Rate	0.1%	0.0%	Improved
Output Throughput	1262 tok/s	1099 tok/s	-12.9%

TTFT (input=8000 tokens, concurrency=1)

Metric	BF16	INT32	Diff
Mean TTFT	382 ms	458 ms	+20%
Mean TPOT	9.54 ms	9.49 ms	-0.5%

Output Stability (request_ali.py, temperature=0)

Metric	BF16	INT32
Valid outputs	5/10 (50%)	10/10 (100%)

Known Limitations

• TTFT ~20% slower due to forced kernel selection (only non-PS non-tkw1 variant modified)
• Throughput ~13% lower for the same reason
• Future improvement: modify all kernel variants (PS, tkw1, tkw1-PS) to restore AITER heuristic

Test Plan

• 10-run bit-identical consistency test
• GSM8K accuracy benchmark
• TTFT latency benchmark
• Output stability test (request_ali.py)