[module_mla_reduce] detorch reduce.cu + delete dead no_redundant Python port

[amd-ruitang3]

Clean-rebuild wall for module_mla_reduce (2 TUs): ~243s (torch-included) → ~222–231s de-torched (~5–9%; device codegen of the 1269-line opus/CK kernel dominates, so torch-header removal is a modest slice)

1. csrc/kernels/mla/reduce.cu — remove torch dependency

The Python side (develop=True), the C++ signature (mla_reduce.h uses aiter_tensor_t) and the pybind (rocm_ops.hpp is torch-free) were already de-torched in #4729; only reduce.cu still pulled torch. Fixed the two remaining torch deps:

• TORCH_WARN (torch macro, host-side occupancy warning) → fprintf(stderr, ...). Same trigger frequency as before (the original used TORCH_WARN, not _ONCE), and it only fires in the degraded high-splits regime — no kernel/GPU perf impact.
• std::optional in the mla_reduce_v1 signature was satisfied only transitively by the torch headers; after dropping <ATen/hip/HIPContext.h> / <ATen/hip/impl/HIPGuardImplMasqueradingAsCUDA.h> / <torch/python.h> it needs an explicit #include <optional> (+ <cstdio> for the fprintf).

2. aiter/ops/attention.py — delete dead code

get_mla_metadata_v1_no_redundant was reimplemented as a pure-Python port in #4729 but has zero in-repo callers and no tests — dead code. Removed it (and its now-unused import heapq).

Why it matters (latent build break)

mla_reduce_v1 is @compile_ops(develop=True), whose production build sets torch_exclude=True — this strips pytorch's -isystem .../torch/include from the compile. The original reduce.cu fails under that path:

fatal error: 'ATen/hip/HIPContext.h' file not found

So this isn't only a build-time win — the de-torched TU is what makes the module actually buildable in the torch-excluded develop path.

Verification (in-container, gfx942 / MI308X)

• op_tests/test_mla_reduce.py — ALL PASSED (all splits / head / dv combos)
• black + ruff check — clean
• Clean-rebuild wall for module_mla_reduce (2 TUs): ~243s (torch-included) → ~222–231s de-torched (~5–9%; device codegen of the 1269-line opus/CK kernel dominates, so torch-header removal is a modest slice)

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4796 --add-label <label>