Upstream sync 13/N: merge 1ab2801dde [ROCm] Remove stale SDPA and skinny GEMM workarounds (#50907) (conflict)

[roberteg16]

Context

Thirteenth step of the batched upstream catch-up. Stacked on #<batch 12 PR>.

Conflict-only step.

Merged upstream changes:

1. 1ab2801dde [ROCm] Remove stale SDPA and skinny GEMM workarounds vllm-project/vllm#50907 — [ROCm] Remove stale SDPA and skinny GEMM workarounds

One conflict, in vllm/model_executor/layers/utils.py, at the wvSplitKrc skinny-GEMM call. Both sides are AMD-authored, which is exactly why this needed reading the kernel rather than trusting either side.

ours:     x_view = x.reshape(-1, x.size(-1))
          with record_function_or_nullcontext(f'wvSplitKrc {n}x{m}x{k}'):
              out = ops.wvSplitKrc(weight, x_view, cu_count, bias)
          return out.reshape(*x.shape[:-1], weight.shape[0])
theirs:   x_view = x.reshape(-1, x.size(-1)).contiguous()
          return ops.wvSplitKrc(x_view, weight, cu_count, bias)
resolved: x_view = x.reshape(-1, x.size(-1)).contiguous()        <- theirs
          with record_function_or_nullcontext(f'wvSplitKrc {n}x{m}x{k}'):
              out = ops.wvSplitKrc(x_view, weight, cu_count, bias)  <- theirs' arg order
          return out.reshape(*x.shape[:-1], weight.shape[0])     <- ours

Three independent differences, resolved separately — two to theirs, one to ours.

Follow-up fix in this branch (f94abd3f06). The weight.is_contiguous()
gate this commit adds to use_skinny disqualifies every weight that
pad_weights_avoid_cache_cliff_on_gfx11 has padded — that pad changes
stride(0) only, on purpose — dropping those layers onto the fallback GEMM.
Measured cost: −3.3% decode on bench2 (83.1 → 80.2 tok/s), bisected to this
batch with the per-PR CI wheels. Relaxed to weight.stride(-1) == 1, which is
what the kernel actually requires: the host captures Kap/Kbp = stride(0) and
the kernel indexes B_[row * Kbp] rather than assuming K. Recovers the full
83.1 tok/s; correctness checked against torch.nn.functional.linear with
padded weights (max |Δ| 0.0039). Details in the
sweep comment.

Audit

1. Argument order — took theirs. This is a real fork bug being fixed.

The two skinny kernels use opposite conventions, so the fork's ordering cannot be justified by analogy with the wvSplitK(weight, x_view, …) call 60 lines below:

csrc/rocm/skinny_gemms.cu:1261  wvSplitK:   M_in = in_a.size(0)   <- weight 1st
csrc/rocm/skinny_gemms.cu:2286  wvSplitKrc: M_in = in_b.size(0)   <- weight 2nd
                                            N_in = in_a.size(0)
                                            K_in = in_b.size(1)

wvSplitKrc reads M and K off its second argument, so the second argument must be the [m, k] weight and the first the activation. The fork passed (weight, x_view), which transposes the problem: M_in would be the token count and K_in the weight's second dim.

2. .contiguous() — took theirs. Upstream's own comment on the sibling call explains it: a shape-preserving reshape can retain a transposed activation's non-contiguous strides, and the skinny kernels assume contiguous K.

3. Profiler label and output reshape — kept ours. The record_function wrapper is on every other GEMM path in this file (DOT, wvSplitK, LLMM1, BLAS — lines 300–348); dropping it here alone would leave one unlabelled path in the profile. The reshape is not redundant either: the kernel allocates out_c as 2-D {N_in, M_in} (skinny_gemms.cu:2308), so for x with more than two dims upstream's bare return hands back a 2-D tensor where the caller expects [*x.shape[:-1], m].

Reachability, stated plainly: this path is gated on on_gfx950(), so it is dead code on gfx1151 (Strix Halo) and the fix cannot be exercised by this fork's benchmarks. Resolved for correctness on gfx950 rather than left to rot; a gfx950 owner should confirm.

py_compile, ruff check and ruff format pass on the merged file.

Merge commit only — do not squash or rebase.

AI assistance was used to prepare this merge.

Test plan

• Argument order verified against the kernel source, not by analogy
• Output-shape contract checked against the kernel's out_c allocation
• py_compile + ruff check + ruff format
• Build + 6-model benchmark sweep vs the Strix Halo dashboard — no regression on transformers 5.15.0 (results)