Annotate moe_sorting_dispatch_policy as int for fused_moe

[nholmber]

Motivation

The type annotation bool was incorrect for moe_sorting_dispatch_policy, which accepts int values. The @torch_compile_guard decorator uses these annotations to generate PyTorch custom op schemas; with bool, PyTorch schema enforcement casts any value to bool, so dispatch_policy=2 becomes bool(2)=True (1), silently losing the intended policy. Using int allows callers to set dispatch_policy=2 correctly.

Fixes: #2576

Technical Details

2-line change in aiter/fused_moe.py:

Line 191: moe_sorting_dispatch_policy: bool = 0 → int = 0 (in fused_moe_fake)
Line 225: moe_sorting_dispatch_policy: bool = 0 → int = 0 (in fused_moe_)

Will be followed up by vLLM change to enable user to set the policy.

Test Plan

• Verified dispatch_policy=2 is passed through correctly to the kernel (previously silently truncated to 1)
• lm_eval gsm8k accuracy on Qwen3-Next-80B-A3B-Instruct-FP8 (MI355X, TP1)
• E2E throughput benchmark (vllm bench serve, 1k/1k, concurrency 16)

Test Result

Metric	dp=0 (baseline)	dp=2 (with fix)
gsm8k flex_extract	0.8560	0.8628
gsm8k strict_match	0.8143	0.8196
Output throughput (tok/s)	1527.83	1551.84 (+1.6%)

No accuracy regression. Modest throughput improvement from correct dispatch policy.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-355	Run Triton tests on MI355 in addition to MI325
ci:sglang	SGLang integration tests
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2639 --add-label <label>