[CK_TILE] Use Unified Workspace for FMHA BWD - #2948
Merged
Merged
Conversation
PR #2321 inlined launcher.dq_acc_splits / needs_zero_dq_acc as a hardcoded `nsplits = ceil(seqlen_k/16)` and unconditional zero, because the CK branch it bumped to had temporarily removed fmha_bwd_launcher. The pinned CK now has the launcher back, so restore the #2216 pattern: construct fmha_bwd_traits + fmha_bwd_launcher and read nsplits and needs_zero_dq_acc from it. Functional behavior is unchanged with the current pinned CK; this is a prep commit to keep the diff for the upcoming #6152 (unified workspace) adaptation small. Touched files: - csrc/py_itfs_ck/mha_bwd_kernels.cu (batch CK entry) - csrc/py_itfs_ck/mha_varlen_bwd_kernels.cu (group CK entry) - op_tests/cpp/mha/benchmark_mha_bwd.cpp (benchmark host)
CK PR #6152 replaces fmha_bwd_args.dq_acc_ptr + the four dq_acc_* strides with a single opaque workspace_ptr, and exposes fmha_bwd_launcher.workspace_size + prepare_workspace(void*) instead of dq_acc_splits + needs_zero_dq_acc. aiter::mha_bwd_args is unchanged (no new fields). The existing dq_acc_ptr field doubles as the CK workspace pointer for the CK path; the four dq_acc_* stride fields are kept for the ASM v3 path that still consumes them. The torch entries (mha_bwd_kernels.cu / mha_varlen_bwd_kernels.cu / benchmark_mha_bwd.cpp) construct the launcher, allocate an at::Tensor workspace of launcher.workspace_size bytes, call launcher.prepare_workspace, and pass workspace.data_ptr() through dq_acc_ptr. For group mode, the launcher requires host-side seqstart arrays. The torch varlen entry copies cu_seqlens_q_padded (or cu_seqlens_q if no padded variant was provided) via .to(at::kCPU) — using the same "physical seqstart" convention that the kernel itself indexes dq_acc with. The aiter::mha_bwd dispatcher does its own hipMemcpy from a.seqstart_q_ptr / a.seqstart_k_ptr for the CK fallback path. Neither host buffer crosses the call boundary. Submodule bumped to 8a59f8afa58 (subtree-split of monorepo users/yiding12/fmha-bwd-workspace tip, includes the per-nhead dq_acc stride fix for group mode).
Removes the dual-purpose dq_acc_ptr field and the four dq_acc_* stride fields from aiter::mha_bwd_args, replacing them with one callback that serves both dispatch paths: std::function<void*(size_t bytes, bool zero_init)> workspace_alloc; - CK fallback (in aiter::mha_bwd): Constructs fmha_bwd_launcher, queries workspace_size, calls workspace_alloc(size, zero_init=false), then forwards the pointer to launcher.prepare_workspace + launcher.run. The torch entries no longer construct fmha_bwd_launcher themselves; py_itfs_ck/* shrinks considerably (traits + launcher + workspace blocks deleted, varlen also drops its private cu_seqlens D2H copy). - ASM v3 (in fmha_v3_bwd): Determines dq_accum shape and byte count internally (based on is_group_mode, v3_atomic_fp32, hdim, batch, nhead, seqlen) and calls workspace_alloc(bytes, zero_init=true). The torch entries no longer allocate or zero dq_accum themselves; py_itfs_cu/asm_mha_*.cu shrinks to a small lambda. The zero_init flag lets each backing storage pick its efficient zero path (torch::zeros, DeviceMem::SetZero, ...) instead of forcing the dispatch path to know HIP memset semantics. ASM kernels need it because they atomically accumulate into dq_accum; the CK launcher fills its own workspace and does not need a pre-zero.
Contributor
🏷️ CI GuideRuns automatically on every PR:
Extended tests (opt-in via labels):
|
Contributor
There was a problem hiding this comment.
Pull request overview
This PR refactors FMHA backward (bwd) call sites to use a unified per-call workspace allocator instead of passing an explicit dq_accum buffer and its layout/strides, aligning CK and ASM paths behind a single “workspace” mechanism.
Changes:
- Add
workspace_alloc(bytes, zero_init)callback tomha_bwd_argsand thread it through Python interfaces and the C++ dispatcher. - Update CK + ASM Python wrappers to allocate a byte workspace tensor on demand (optionally zero-initialized).
- Update the C++ benchmark harness to pre-allocate a single device buffer sized to satisfy both ASM dq_accum and CK workspace needs.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
csrc/include/mha_bwd.h |
Adds workspace_alloc callback to mha_bwd_args and removes explicit dq_acc fields/strides. |
csrc/cpp_itfs/mha_bwd.cu |
Uses workspace_alloc for CK launcher workspace and ASM dq_accum allocation; adds host-side seqstart handling for group mode. |
csrc/py_itfs_ck/mha_bwd_kernels.cu |
Switches from dedicated dq_accum tensor to workspace_alloc for CK bwd wrapper. |
csrc/py_itfs_ck/mha_varlen_bwd_kernels.cu |
Same as above for varlen CK bwd wrapper. |
csrc/py_itfs_cu/asm_mha_bwd.cu |
Switches from dedicated dq_accum tensor to workspace_alloc for ASM bwd wrapper. |
csrc/py_itfs_cu/asm_mha_varlen_bwd.cu |
Same as above for varlen ASM bwd wrapper. |
op_tests/cpp/mha/benchmark_mha_bwd.cpp |
Pre-allocates a single device buffer and serves it via workspace_alloc to satisfy both backends. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
- mha_bwd CK fallback: explicitly reject group mode with missing seqstart pointers (LOG_ERROR + return -1) instead of silently passing nullptr to the launcher. Matches the existing AITER_LOG_WARNING pattern in fmha_v3_bwd for unsupported configurations. - benchmark workspace_alloc: replace silent return-nullptr-on-oversize with AITER_CHECK so the failure is loud at the actual call site rather than as an opaque kernel crash later. Also tighten the zero-init path to hipMemset only `bytes` rather than the full pre-allocated buffer.
slippedJim
previously approved these changes
May 8, 2026
1 task
2 tasks
valarLip
approved these changes
May 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Adapts AITER to ROCm/rocm-libraries#6152 ("CK_TILE Use Unified Workspace for FMHA BWD"), which replaces the per-arg
dq_acc_ptr+ 4 stride fields infmha_bwd_argswith a single opaqueworkspace_ptrand exposesfmha_bwd_launcher.workspace_size/prepare_workspace(void*).Technical Details
Cleans up
aiter::mha_bwd_argsABI: dropsdq_acc_ptrand the 4 dq_acc strides, replacing them with one callback that serves both dispatch paths:aiter::mha_bwd): constructsfmha_bwd_launcher, queriesworkspace_size, callsworkspace_alloc(size, false), forwards the pointer toprepare_workspace+launcher.run. Torch entries no longer construct the launcher themselves;py_itfs_ck/*shrinks (traits + launcher + workspace blocks deleted, varlen also drops its privatecu_seqlensD2H copy).fmha_v3_bwd): determinesdq_accumshape / byte count internally (based onis_group_mode,v3_atomic_fp32,hdim,batch,nhead,seqlen) and callsworkspace_alloc(bytes, true). Torch entries no longer allocate or zerodq_accumthemselves;py_itfs_cu/asm_mha_*.cushrinks to a small lambda.The
zero_initflag lets each backing storage pick its efficient zero path (torch::zeros,DeviceMem::SetZero, ...) instead of forcing the dispatch path to know HIP memset semantics. ASM kernels need zero-init because they atomically accumulate intodq_accum; the CK launcher fills its own workspace viaprepare_workspace.Test Plan
On gfx950 inside
rocm/pytorch:rocm7.2_ubuntu24.04_py3.12_pytorch_release_2.9.1:python op_tests/test_mha.pypython op_tests/test_mha_varlen.pyIn addition, we have CK side downstream passed: jenkins / rocm-libraries-folder/Composable Kernel / PR-6152 / #14
Test Result
All cases pass. dQ / dK / dV
max diff≤ 0.008 (fp16 numerical tolerance ~0.01); noAssertionError/Traceback/hipError.Submission Checklist