mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-11 00:39:02 +00:00
e7798e9560b5cb9079907afa472bf002e2993755
3314 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
JP-Fernando
|
e7798e9560 |
[rocm-libraries] ROCm/rocm-libraries#7112 (commit a6e5eac)
Add asynchronous XOR shuffle support to the Async GEMM pipeline and the MX GEMM pipeline (#7112) ## Motivation The goal of this work is to apply XOR shuffle (swizzle) to the current `comp_async` GEMM pipeline and the `gemm_mx` pipeline. XOR swizzling has been helpful to avoid LDS bank conflicts, as data are redistributed across LDS banks, such that simultaneous threads accessing different rows land on different LDS banks. ## Technical Details A similar approach to the work in the existing eight-waves pipeline was followed. Currently, XOR swizzle support is available for FP8 and BF8 types. FP4 support is also available for MX GEMM. Should the types not match, or should the async vector width be of an unsupported size, then the pipeline falls through to the previously existing ('unswizzled') path. ## Test Plan Execute `test_ck_tile_gemm_pipeline_comp_async` for the Async GEMM pipeline. Execute `test_ck_tile_mx_gemm_fp8` and `test_ck_tile_mx_gemm_fp4` for the MX GEMM pipeline. ## Test Result The tests passed successfully in the `Alola` cluster with MI350 hardware. ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Fernando Jiménez <fernando.jimenez@streamhpc.com> Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com> |
||
|
Thrupti Raj Lakshmana Gowda
|
c31fc4df52 |
[rocm-libraries] ROCm/rocm-libraries#7311 (commit 79d8cae)
[CK Tile Engine] Daily tier sampling for tile engine GEMM (#7311) Summary - Replace uniform random instance sampling (random.shuffle) with scrambled Sobol + Latin Hypercube + maximin space-filling sampling, per the Tile Engine Benchmark Sampling RFC - Add op-weighted budget allocation via new TILE_ENGINE_SAMPLING_TIER=daily CMake knob that auto-distributes 8,000 instances across ops proportional to registered weights in op_weights.json - Emit chosen_instances.json manifests for reproducibility tracking - Consolidate 5 copies of sampling logic into single _apply_sampling() method on the base class Jenkinsfile changes Replace per-op -D *_MAX_INSTANCES=250 with single -D TILE_ENGINE_SAMPLING_TIER=daily in gfx942/gfx950/gfx1201 stages. Budget auto-distributes (8000 total per GPU target). --------- Co-authored-by: Claude Sonnet 4 <noreply@anthropic.com> |
||
|
Wojciech Laskowski
|
275629fe34 |
[rocm-libraries] ROCm/rocm-libraries#6014 (commit 2f8259d)
[CK Tile] Adding MFMA wrappers for dense builtins (#6014) ## Motivation This PR is part of the [WMMA/MFMA] unification work. It's the second of the series of PRs (after #5801) that add all the necessary MMA builtins as `amdgcn_mma` structs. This PR focuses on dense MFMA intrinsics. ## Technical Details This change adds new specializations for WMMA dense builtins. In total, we add 55 MFMA builtins. ## Test Plan All the new wrappers were added to the test suite in `test_amdgcn_mma_layout.inc`. ## Test Result Test pass locally, waiting for the CI. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
dependabot[bot]
|
cad4f236d4 |
[rocm-libraries] ROCm/rocm-libraries#7603 (commit 4ea31a8)
Bump idna from 3.11 to 3.15 in /projects/composablekernel/docs/sphinx (#7603) Bumps [idna](https://github.com/kjd/idna) from 3.11 to 3.15. <details> <summary>Changelog</summary> <p><em>Sourced from <a href="https://github.com/kjd/idna/blob/master/HISTORY.md">idna's changelog</a>.</em></p> <blockquote> <h2>3.15 (2026-05-12)</h2> <ul> <li>Enforce DNS-length cap on individual labels early in <code>check_label</code>, short-circuiting contextual-rule processing for oversized input while staying compatible with UTS 46 usage.</li> <li>Tidy core helpers: hoist bidi category sets to module-level frozensets (avoiding per-codepoint list construction), simplify length checks, and reuse the shared <code>_unicode_dots_re</code> from <code>idna.core</code> in the codec module.</li> <li>Use <code>raise ... from err</code> for proper exception chaining and switch internal string formatting to f-strings.</li> <li>Allow <code>flit_core</code> 4.x in the build backend.</li> <li>Expand the ruff lint set (flake8-bugbear, flake8-simplify, pyupgrade, perflint) and apply the surfaced fixes; pin lint CI to Python 3.14.</li> <li>Add Dependabot configuration for GitHub Actions.</li> <li>Convert README and HISTORY from reStructuredText to Markdown.</li> <li>Reference CVE-2026-45409 for the 3.14 advisory in place of the initial GHSA identifier.</li> </ul> <p>Thanks to Felix Yan, Stan Ulbrych, and metsw24-max for contributions to this release.</p> <h2>3.14 (2026-05-10)</h2> <ul> <li>Removed opportunity to process long inputs into quadratic time by rejecting oversize inputs up-front. Closes a bypass of the CVE-2024-3651 mitigation. [CVE-2026-45409]</li> </ul> <p>Thanks to Stan Ulbrych for reporting the issue.</p> <h2>3.13 (2026-04-22)</h2> <ul> <li>Correct classification error for codepoint U+A7F1</li> </ul> <h2>3.12 (2026-04-21)</h2> <ul> <li>Update to Unicode 17.0.0.</li> <li>Issue a deprecation warning for the transitional argument.</li> <li>Added lazy-loading to provide some performance improvements.</li> <li>Removed vestiges of code related to Python 2 support, including segmentation of data structures specific to Jython.</li> </ul> <p>Thanks to Rodrigo Nogueira for contributions to this release.</p> </blockquote> </details> <details> <summary>Commits</summary> <ul> <li><a href=" |
||
|
JH-Leon-KIM-AMD
|
720ceb6500 |
[rocm-libraries] ROCm/rocm-libraries#7528 (commit b4cae6f)
[CK Tile] Support multi-vector reads in static encoding patterns (#7528) ## Motivation The thread-raked / warp-raked / block-raked static tile distribution patterns in `ck_tile` silently produce wrong results when the contiguous tile dimension is larger than `warp_size * vector_size`, because the encoding has no per-thread iteration dimension along X. Concretely, with `M_Tile=N_Tile=128`, `VectorSize{A,B,C}=1` in `ConvConfigComputeV3`, the grouped convolution backward-weight example reports about 50 percent wrong values, with errors starting exactly at the `X0*X1 = 64` boundary. The second pass over the contiguous dim is never performed. This PR extends the encoding so multi-vector reads in the contiguous tile dimension are supported, while keeping every existing call site bit-for-bit identical. ## Technical Details Three files changed. ### 1. `include/ck_tile/core/algorithm/static_encoding_pattern.hpp` Add a per-thread X iteration dimension in all three raked specializations: - `X0 = min(warp_size, XPerTile / X1)` — threads in X dim - `X1 = min(LargestVec, VecSize)` — vector size per access - `X2 = XPerTile / (X0 * X1)` — number of X-iters per thread (new) `X2` is gated with `if constexpr (X2 == 1) { old } else { new }` in both `make_2d_static_tile_distribution()` and `make_shuffled_2d_static_tile_distribution()`. The new encoding places `X2` in the middle of the Ys iteration list, which preserves reverse symmetry between the regular `<..., X2, X1>` and shuffled `<X1, X2, ...>` encodings. Patterns updated: `thread_raked`, `warp_raked`, `block_raked`. ### 2. `include/ck_tile/core/tensor/transpose_tile.hpp` Added a parallel `else if constexpr (... && NDimY == 3 && ...)` branch alongside the existing `NDimY == 2` branch. The original branch is byte-for-byte unchanged. Both branches dispatch to the same `transpose_tile2d_impl_in_thread`, whose body has always been NDimY-generic (iterates with `static_for<0, NDimY, 1>` and `number<NDimY>{}`). ### 3. `experimental/grouped_convolution_tile_instances/generate_instances.py` Removed the two now-obsolete skip guards in `parse_bwd_weight_instances` and `parse_bwd_data_instances`: ```python if m_per_block > (warp_size * a_scalar_per_vector) or n_per_block > (warp_size * b_scalar_per_vector): print(f"Skipping instance {instance_id} with multiple warps per continous tile dim since it's not supported yet.") continue ``` Other unrelated skips (V5 / V6 / ASYNC_V4 pipeline gating, irregular-load shapes, scalar-per-vector > tile size) are kept untouched. ### Compatibility Strict. Every existing caller has `X2 == 1` and therefore hits the original encoding path verbatim. No upstream config or pipeline behavior changes. ## Test Plan The grouped convolution example is the natural exerciser since `GroupedConvUniversalPipelineAgBgCrPolicy` selects `thread_raked` for both A and B tiles, and all three conv directions share the same `ConvConfigComputeV3`. For each test below we ran: ``` ./build/bin/tile_example_grouped_conv_bwd_weight [-prec={fp16,bf16}] ./build/bin/tile_example_grouped_conv_fwd [-prec={fp16,bf16}] ./build/bin/tile_example_grouped_conv_bwd_data [-prec={fp16,bf16}] ``` with `ConvConfigComputeV3` tile/vector parameters tweaked to cover both code paths: | Test | M / N / K | VecA/B/C | A path | B path | dtype | |------|-------------|----------|------------|----------------|-------------| | T1 | 16/64/32 | 4/8/4 | old (X2=1) | old (X2=1) | fp16 | | T2 | 128/128/64 | 2/2/2 | old (X2=1) | old (X2=1) | fp16 | | T3 | 256/256/64 | 1/1/1 | old (X2=1) | new (X2=4) | fp16 | | T5 | 256/256/64 | 1/1/1 | old (X2=1) | new (X2=4) | fp16 (3 dir)| | T4b | 128/128/128 | 1/1/1 | new (X2=2) | new (X2=2) | fp16 + bf16 (3 dir) | A larger T4a (256/256/128) was attempted to stress both A and B with X2>1 on bigger tiles but was blocked by the gfx942 hardware LDS cap (128 KB > 64 KB limit), independent of this PR. For the generator change we ran: ``` python3 generate_instances.py --mode profiler --direction all ``` and verified `Skipping instance ... with multiple warps per continous tile dim` no longer appears (count went from non-zero to 0); other skip categories are unchanged. `clang-format-18` was applied to both modified `.hpp` files (matches the repo's `.clang-format`). ## Test Result - T1 and T2 (compat-strict, every X2 is 1, old code path): `correct`. Confirms existing callers are unaffected. - T3 (X2=4 on B only): `correct`. First true exercise of the new NDimY=3 encoding + transpose branch. - T5 (T3 across `fwd` + `bwd_data` + `bwd_weight`, fp16): all 3 `correct`. - T4b (X2>1 on both A and B, fp16 + bf16, all 3 directions): all 6 runs `correct`. - Generator: 0 `multiple warps per continous tile dim` skips remaining; other skips unchanged. Sample run output (T4b, bf16, bwd_data): ``` shape: tile_gemm_shape_128x128x128x4_1x4x1_16x16x32 pipeline: pipeline_AgBgCrCompV3_128x128x128_256_1x1x1_1x4_1x1x1_..._DoubleSmemBuffer_0 Vector size A: 1, Vector size B: 1, Vector size C: 1 0.934907 ms, 8.34683 TFlops, 34.3178 GB/s Relative error threshold: 0.00390625 Absolute error threshold: 0.25 The CPU verification result is: correct ``` ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Kiefer van Teutem
|
b5f8bef97f |
[rocm-libraries] ROCm/rocm-libraries#6088 (commit 6ac353c)
[CK Tile][MFMA/WMMA unification] Add support for packed datatypes (tiny types) (#6088) ## Motivation This MR makes all the changes required for the unified architecture to be able to deal with packed datatypes i.e. int4, fp4, fp6, and bf6. The crux is that layout parameters should be interpreted as describing the pure mathematical matrix fragments, while the ext_vectors and tile distribution encodings describe everything in terms of packed datatype units. This matches how packed types are dealt with in ck_tile and should play nicely with the load and store tile ops once we integrate the unified framework into CK tile. The bf6 datatype was added to CK tile in the form of pk_bf6x16_t and pk_bf6x32_t, which did not exist before. The ext_vector implementations of pk_fp6x16_t and pk_bf6x16_t (vec size 1 and 2) were extended to make the subscripting operator work as expected. The layout test was adapted to be compatible with all packed datatypes, and all new intrinsics were added to the test. This MR adds ALL intrinsics across ALL architectures which use packed datatypes, as well as ALL scale intrinsics: mfma_scale_f32_16x16x128_f8f6f4 gfx950 (F8xF8, BF8xBF8, F4xF4, F6xF6, BF6xBF6) mfma_scale_f32_32x32x64_f8f6f4 gfx950 (F8xF8, BF8xBF8, F4xF4, F6xF6, BF6xBF6) wmma_i32_16x16x16_iu4_w32 wmma_i32_16x16x16_iu4_w32_gfx12 wmma_i32_16x16x32_iu4_w32_gfx12 ## Testing All intrinsics were tested on all architectures. |
||
|
Aviral Goel
|
458dd0ac4c |
[rocm-libraries] ROCm/rocm-libraries#7130 (commit 9e1e065)
[CK_TILE] Redesign LDS store API with pre-computed window coordinates (+15% MI355X, +6% MI300X) (#7130) ## Summary - Redesign the LDS store API to separate window creation from memory transfer - Add `MakeDistributedLdsStoreWindow` factory, `LocalStore` (fast path), and `LocalStoreWithCoordRecompute` (slow path) to the pipeline base class - Convert CompV3 as the reference implementation - Document the slow/fast path distinction across core tensor headers ## Motivation `LocalPrefill` hides a performance cliff: when given a bare `tile_window_with_static_lengths`, it silently reconstructs `tile_window_with_static_distribution` on every call — paying significant VALU overhead (~96 for typical configurations) for XOR coordinate computation. The cost is invisible at the call site. The new API makes the cost explicit via three verbs: | Verb | Method | Cost | When to use | |------|--------|------|-------------| | **Create** | `MakeDistributedLdsStoreWindow(bare, dstr)` | VALU (once) | Before hot loop, when VGPR budget allows | | **Store (fast)** | `LocalStore(precomputed_window, tensor)` | 0 VALU for coords | Pre-computed window available | | **Store (on-the-fly)** | `LocalStoreWithCoordRecompute(bare, tensor)` | VALU per call | VGPR budget tight, or one-shot stores | Both `LocalStore` and `LocalStoreWithCoordRecompute` enforce correct window types via `static_assert`. `LocalPrefill` is retained for backward compatibility (69 call sites across 6 pipeline files). ## Performance ### 86 Shapes, CompV3_2 (128×128 tile), fp16, RCR layout **gfx942 (MI300X): 86/86 improved, 0 regressions. Average gain: +6.2%** **gfx950 (MI355X): 85/86 improved, 1 neutral, 0 regressions. Average gain: ~+15%** <img width="2777" height="1178" alt="pr7130_perf_chart" src="https://github.com/user-attachments/assets/b2f5c406-eb20-469d-8da6-dd608c28fbcc" /> | Shape (MxNxK) | Source | gfx942 | gfx950 | |---|---|---|---| | 22016x256x4096 | llama2_7b_fc1 | +5.3% | +11.4% | | 22016x512x4096 | llama2_7b_pfill | +5.9% | +10.9% | | 4096x512x22016 | llama2_7b_pfill | +7.6% | +28.5% | | 22016x1024x4096 | llama2_7b_pfill | +6.1% | +10.1% | | 4096x1024x22016 | llama2_7b_pfill | +7.4% | +17.2% | | 22016x4096x4096 | llama2_7b_pfill | +5.2% | +9.3% | | 4096x4096x22016 | llama2_7b_pfill | +6.0% | +9.3% | | 4096x4096x4096 | llama2_7b_pfill | +5.7% | +10.6% | | 28672x256x4096 | llama3_8b_fc1 | +5.4% | +12.2% | | 28672x512x4096 | llama3_8b_pfill | +4.9% | +6.4% | | 4096x512x28672 | llama3_8b_pfill | +7.4% | +1.5% | | 28672x2048x4096 | llama3_8b_pfill | +4.9% | +8.6% | | 4096x2048x28672 | llama3_8b_pfill | +6.4% | +8.4% | | 28672x8192x4096 | llama3_8b_pfill | +5.4% | +8.0% | | 7168x1024x8192 | llama70b_pfill | +6.6% | +10.8% | | 8192x1024x7168 | llama70b_pfill | +6.4% | +11.4% | | 7168x4096x8192 | llama70b_pfill | +6.2% | +9.6% | | 16384x256x4096 | bloom_fc1 | +6.4% | +20.3% | | 16384x512x4096 | bloom_fc1 | +5.8% | +8.5% | | 16384x1024x4096 | bloom_fc1 | +6.0% | +10.9% | | 16384x2048x4096 | bloom_fc1 | +5.3% | +10.1% | | 16384x3072x4096 | bloom_fc1 | +5.5% | +8.8% | | 16384x4096x4096 | bloom_fc1 | +5.7% | +8.8% | | 4096x256x16384 | bloom_fc2 | +7.8% | +33.6% | | 4096x512x16384 | bloom_fc2 | +7.5% | +31.6% | | 4096x1024x16384 | bloom_fc2 | +7.1% | +17.1% | | 4096x2048x16384 | bloom_fc2 | +6.9% | +11.0% | | 4096x3072x16384 | bloom_fc2 | +6.8% | +11.0% | | 4096x4096x16384 | bloom_fc2 | +6.7% | +10.3% | | 12288x256x4096 | bloom_inproj | +6.7% | +22.0% | | 12288x512x4096 | bloom_inproj | +6.2% | +9.8% | | 12288x1024x4096 | bloom_inproj | +5.9% | +12.4% | | 12288x2048x4096 | bloom_inproj | +5.8% | +10.1% | | 12288x3072x4096 | bloom_inproj | +5.4% | +10.1% | | 12288x4096x4096 | bloom_inproj | +5.7% | +9.1% | | 250880x256x4096 | bloom_logits | +2.6% | +0.5% | | 4096x256x4096 | bloom_outproj | +7.1% | +28.4% | | 4096x512x4096 | bloom_outproj | +6.8% | +27.4% | | 4096x1024x4096 | bloom_outproj | +6.5% | +21.3% | | 4096x2048x4096 | bloom_outproj | +5.9% | +13.1% | | 4096x3072x4096 | bloom_outproj | +5.9% | +12.0% | | 16x1536x7168 | deepseek | +7.7% | +34.7% | | 32x1536x7168 | deepseek | +7.7% | +34.9% | | 64x1536x7168 | deepseek | +7.6% | +31.3% | | 128x1536x7168 | deepseek | +7.6% | +25.8% | | 256x1536x7168 | deepseek | +7.7% | +27.9% | | 512x1536x7168 | deepseek | +7.6% | +29.1% | | 1024x1536x7168 | deepseek | +7.3% | +28.8% | | 2048x1536x7168 | deepseek | +6.9% | +20.5% | | 4096x1536x7168 | deepseek | +6.3% | +11.0% | | 8192x1536x7168 | deepseek | +6.2% | +11.3% | | 16384x1536x7168 | deepseek | +6.0% | +9.1% | | 20480x1536x7168 | deepseek | +4.8% | +9.3% | | 16x3072x1536 | deepseek | +6.3% | +25.1% | | 32x3072x1536 | deepseek | +6.4% | +25.3% | | 64x3072x1536 | deepseek | +6.4% | +24.8% | | 1024x1024x1024 | square | +5.5% | +18.7% | | 2048x2048x2048 | square | +6.0% | +19.2% | | 3584x3584x3584 | square | +5.3% | +11.2% | | 5120x5120x5120 | square | +6.1% | +10.0% | | 6144x6144x6144 | square | +5.5% | +9.8% | | 8192x8192x8192 | square | +6.0% | +8.2% | | 1024x4608x1024 | midsize | +4.6% | +4.6% | | 512x18432x512 | midsize | +1.9% | +10.1% | | 4096x18432x4096 | midsize | +5.8% | +8.8% | | 320x8192x320 | stablediff | +4.0% | +11.3% | | 640x2048x640 | stablediff | +4.5% | +14.0% | | 320x8192x1280 | stablediff | +5.6% | +20.1% | | 1x1280x8192 | skinny_m1 | +7.7% | +35.3% | | 1x8192x1024 | skinny_m1 | +6.0% | +20.3% | | 1x7168x8192 | skinny_m1 | +7.7% | +36.6% | | 1x8192x3584 | skinny_m1 | +7.3% | +27.9% | | 1x13312x6656 | skinny_m1 | +7.6% | +30.3% | | 1x13312x16384 | skinny_m1 | +7.8% | +4.2% | | 1x16384x6656 | skinny_m1 | +7.5% | +28.7% | | 1x16384x16384 | skinny_m1 | +7.7% | +2.3% | | 16x4096x4096 | skinny_m16 | +7.4% | +31.9% | | 16x22016x4096 | skinny_m16 | +7.5% | +26.5% | | 16x28672x4096 | skinny_m16 | +7.0% | +15.1% | | 16384x1280x8192 | skinny_m16 | +5.6% | +8.7% | | 16384x8192x1024 | skinny_m16 | +4.5% | +8.8% | | 2048x4096x2048 | mixed | +4.7% | +9.0% | | 4096x2048x8192 | mixed | +6.8% | +11.0% | | 8192x4096x4096 | mixed | +5.2% | +10.0% | | 1x4096x4096 | mixed | +7.4% | +32.4% | | 1024x1024x4096 | mixed | +7.1% | +27.4% | ### ISA Hot Loop Diff (LBB1_32, per K-iteration, gfx942) | Metric | Baseline | Optimized | Delta | |--------|----------|-----------|-------| | Total VALU | 621 | 500 | **-121** | | VGPR / SGPR | 512 / 96 | 512 / 96 | unchanged | ### Hardware Counters — Instruction Mix (gfx950, rocprofiler-compute) Profiled on MI350X, shape 4096×256×16384 (bloom_fc2). Instruction counts are deterministic hardware counters. | Metric | Baseline | Optimized | Δ | |--------|----------|-----------|---| | **VALU instructions/kernel** | 4,642,473 | 987,958 | **−78.7%** | | **INT32 VALU** | 2,592,786 | 541,129 | **−79.1%** | | Instructions / wavefront | 39,178 | 24,400 | −37.7% | | VGPRs (avg) | 98 | 90 | −8% | | **MFMA instructions** | 2,059,702 | 2,059,702 | **0%** | | **LDS instructions** | 1,564,891 | 1,564,891 | **0%** | | **VMEM instructions** | 520,996 | 520,996 | **0%** | MFMA as fraction of total instructions: **30.7% → 67.5%**. Eliminating ~3.65M redundant INT32 VALU instructions (XOR coordinate recomputation per K-iteration) leaves the scheduler more headroom for MFMA dispatch, directly explaining the benchmark gains. |
||
|
Enrico Degregori
|
9565ca21ec |
[rocm-libraries] ROCm/rocm-libraries#5552 (commit 369c7a2)
[CK Tile] Eight Waves pipeline for MX GEMM (#5552) ## Motivation Integrate Eight Waves pipeline in MX GEMM ## Technical Details - EightWaves pipeline: - Add pipeline, policy and block gemm (internally using existing implementation used by GEMM and ABQuant) - Extend support of EightWaves policy for FP4 (packed types) - Async pipeline: - Fix pipeline with packed scales (requires MRepeat and NRepeat to be contiguous) - block gemm specific for MX GEMM is defined because distribution encodings have changed - CShuffle: - Add new functionality to support MRepeat and NRepeat contiguous (defined by `TilesPacked`) - Examples: - Refactor examples to easily switch different configurations (similar to GEMM universal) - Scales values generated consistently with other microscale implementations in CK Tile - Add configuration for EightWaves pipeline - Tests: - Unify existing FP8 and FP4 tests - Add tests for EightWaves pipeline - Scales values generated consistently with other microscale implementations in CK Tile Note: FP6 support for MX GEMM was added later and the support for the Eight Waves pipeline will be done in following PR ## Test Plan Add new pipeline to tests: `test_ck_tile_mx_gemm_async` for both FP4 and FP8 ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Illia Silin
|
f01a8cb28d |
[rocm-libraries] ROCm/rocm-libraries#7547 (commit 7e032ad)
[CK] fix daily builds for pytorch (#7547) ## Motivation This will restore the daily builds that test whether the latest pytorch code can build with the latest CK code (pulled from the standalone CK repo). ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Aaryaman Vasishta
|
fad83d9c90 |
[rocm-libraries] ROCm/rocm-libraries#7016 (commit 2b73c00)
[CK] Fix RDNA3 FMHA tile-load paths (#7016) ## Summary Fix CK tile FMHA paths needed for RDNA3/RDNA4 targets. ## Details This PR addresses RDNA-specific issues hit while enabling xFormers CK FMHA on gfx11/gfx12: - On RDNA3, update FMHA P tile handling so the layout consumed by the second GEMM matches the WMMA path. ## Testing Validated downstream with xFormers CK/FMHA on gfx1201/gfx1151. ```text pytest --import-mode=importlib -q \ tests/test_mem_eff_attention.py::test_forward \ tests/test_mem_eff_attention.py::test_backward \ tests/test_mem_eff_attention.py::test_dropout_ck 3844 passed, 5244 skipped, 26 warnings --------- Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com> |
||
|
Yung-sheng Tu
|
5169cd14a1 |
[rocm-libraries] ROCm/rocm-libraries#7543 (commit 2b735ff)
Fix for #6207 (#7543) ## Motivation PR #6207 introduces an error. This PR is the fix of it. ## Technical Details Adds a path for GFX1250 in `to_string` ## Test Plan Test has already included. ## Test Result Test should pass. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Po Yen Chen
|
a11f53564f |
[rocm-libraries] ROCm/rocm-libraries#7530 (commit 378e049)
[CK] Fix FMHA sink dispatch when init_sink_value is set (#7530) ## Summary - Fix `traits.has_sink` in `fmha_fwd_runner.hpp` to also check `init_sink_value != 0`, so the GPU kernel dispatches with sink support when `-init_sink=1` is passed. - Gate `run_sink_mask_tests` (StreamLLM) and `run_sink_init_tests` (GPT-OSS) behind opt-in flags `-m` and `-g` in `smoke_test_fwd.sh`. These tests require sink=true kernel instances which are excluded by the `BUILD_TESTING` CMake filter (`*_nsink*`), causing unconditional "not supported yet" failures (48 tests in CI). The opt-in flag approach was borrowed from PR #6057. ## Why gate tests instead of compiling sink=true kernels? The `BUILD_TESTING` filter in `CMakeLists.txt` uses `*_nsink*` glob patterns for the `fwd` and `fwd_splitkv` APIs, excluding sink=true kernel instances from compilation. We chose opt-in flags over widening the filter because: - **Compile time**: Enabling sink=true kernels doubles the kernel variants for `fwd` and `fwd_splitkv` APIs. The filter exists specifically to reduce CI build times. - **Incremental enablement**: Sink support (StreamLLM / GPT-OSS) is still maturing. Gating lets teams opt in explicitly (`smoke_test_fwd.sh -g`) while keeping the default CI path fast. - **Precedent**: splitkv (`-s`) and appendkv (`-a`) tests already follow this opt-in pattern. ## Test plan - [ ] Run `smoke_test_fwd.sh -g` with sink=true kernels compiled and verify sink-enabled kernels are dispatched - [ ] Verify `smoke_test_fwd.sh` still passes without `-m` / `-g` flags - [ ] Confirm CI no longer fails on sink tests (they are now opt-in) |
||
|
Johannes Graner
|
3727d5220a |
[rocm-libraries] ROCm/rocm-libraries#5652 (commit 7dc7d1d)
[CK Conv] Wavelet gemm pipeline for bwd_weight convolution (#5652) ## Motivation In the current CShuffleV3 backward weight kernel, the in-kernel conv-to-GEMM transform generates significant INT32 VALU pressure per MFMA instruction. On VALU-heavy shapes (e.g., G=1, 3×3, C=256), these index computation ops compete with MFMA for VALU issue slots, creating a bottleneck that cannot be resolved by pipeline prefetching alone. This PR adds a wave-specialized ("wavelet") convolution backward weight kernel that splits workgroup threads into two roles: - **Load waves**: conv-to-GEMM address computation + global memory loads + LDS writes (all VALU/VMEM) - **Math waves**: LDS reads + MFMA + CShuffle epilogue (no index computation) By physically separating the two instruction classes onto different waves, VALU and MFMA execute on different hardware functional units without contention. ## Technical Details **Core kernel (new files):** - `gridwise_gemm_xdl_waveletmodel_cshuffle_conv_v3.hpp` — wave-specialized gridwise GEMM for conv bwd weight (2-way split: load + math) - `device_grouped_conv_bwd_weight_xdl_waveletmodel_cshuffle_v3.hpp` — device op following CShuffleV3 patterns; `BlockSize = TileMathThreadGroupSize` for MFMA wave assignment, `LaunchBlockSize = TileLoad + TileMath` for kernel launch **Wave pipeline (modified):** - `gridwise_gemm_waveletmodel.hpp` — load/math wave pipeline structs with `sched_group_barrier` scheduling hints to front-load VMEM reads before address-advance VALU **Two wave ratios:** - **(4,4)**: 256 load + 256 math = 512 threads (8 waves). Best on large shapes. - **(4,2)**: 256 load + 128 math = 384 threads (6 waves). Best on small shapes (fewer sync barriers, denser MFMA per math wave). **Instance coverage (F16 and BF16 symmetric):** | Ratio | Tiles | Layouts | ConvSpecs | |-------|-------|---------|-----------| | (4,4) | M128×N128, M64×N64, M128×N64, M64×N128 | 2D NHWGC, 3D NDHWGC | Default, Filter1x1Stride1Pad0 | | (4,2) | M64×N64, M128×N64, M64×N128 | 2D NHWGC | Default, Filter1x1Stride1Pad0 | **Existing wavelet model fixes:** - `BlockSize` corrected from `math::max(TileLoad, TileMath)` to `TileMathThreadGroupSize` in the flat-GEMM wavelet device op and gridwise kernel ## Test Plan - `test_grouped_convnd_bwd_weight` GTest: 34 hardcoded test cases covering 1D/2D/3D, F16/BF16, G=1/2/16, various spatial sizes - Performance benchmark: all 37 RetinaNet bwd_weight shapes on gfx950 ```bash ninja -C build test_grouped_convnd_bwd_weight ./build/bin/test_grouped_convnd_bwd_weight ``` ## Test Result **Correctness:** 34/34 GTest cases passed (F16/BF16 × 1D/2D/3D × Default/Filter1x1Stride1Pad0 × various G/N/K/C combinations). **Performance:** Wavelet is the fastest overall instance on 12/37 RetinaNet shapes — all G=1, 3×3 convolutions with C=256 (the VALU-heavy target shapes): | Shape | Uplift vs best baseline | |-------|------------------------| | K=36, 7×7 | 1.91x | | K=36, 100×100 | 1.60x | | K=36, 13×13 | 1.43x | | K=36, 25×25 | 1.38x | | K=36, 50×50 | 1.38x | | K=256, 100×100 | 1.24x | | K=256, 13×13, s=2 | 1.20x | | K=256, 25×25, s=2 | 1.20x | | K=256, 7×7 | 1.17x | | K=256, 13×13 | 1.13x | | K=2376, 50×50 | 1.05x | | K=2376, 100×100 | 1.06x | Where wavelet does not win (25/37): 1×1 convolutions (explicit kernel does host-side transform), grouped convolutions with small per-group channels, and shapes where standard CShuffleV3 already amortizes VALU overhead. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: jakpiase <jakpia21@gmail.com> |
||
|
Bartłomiej Kocot
|
add7627a40 |
[rocm-libraries] ROCm/rocm-libraries#7021 (commit 0766457)
[CK][CK Tile] Grouped Conv Tile Profiler Verification (#7021) ## Motivation Improve CK Tile Conv Profiler for perf measurements. ## Technical Details Add option to disable verification and script to run conv tile profiler. ## Test Plan CI ## Test Result Pending ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. AICK-75 |
||
|
JH-Leon-KIM-AMD
|
9a5d1ea791 |
[rocm-libraries] ROCm/rocm-libraries#6208 (commit 33424f6)
[CK] Enable grouped conv bwd data to match non-grouped perf via NoShuffle + packed descriptors (#6208) ## Motivation Improve performance of grouped convolution backward-data kernels to match non-grouped kernel performance for G=1 cases. ## Technical Details - Add NoShuffle epilogue path (direct VGPR→Global writes) by setting `CDEBlockTransferScalarPerVector_NPerBlock = 1` - Add nongrouped-match instances with optimized BBlockTransfer parameters for better thread utilization - Add packed (flat) descriptor path for G=1 2D convolutions, using simpler tensor descriptors with fewer transform layers to reduce address computation overhead in the GEMM main loop - Cherry-pick PR #6090 for fair benchmarking (cache flush, include dX zeroing cost) ## Test Plan - Benchmark grouped vs non-grouped kernels on MI300X (589 shapes, BF16) - Verify correctness with existing conv bwd data tests ## Test Result | Metric | Before | After | |--------|--------|-------| | Mean ratio (grouped/nongrouped) | 1.159 | **1.028** | | Median ratio | 1.142 | **1.026** | | Cases within 2% | 26 (4.4%) | **186 (31.8%)** | | Cases >20% slower | 188 (32%) | **2 (0.3%)** | NoShuffle + nongrouped-match instances achieve **~2.8% average gap** with non-grouped kernels (down from ~16%). ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: root <root@ctr-cx64-mi300x-4.amd.com> Co-authored-by: root <root@ctr-cx71-mi300x-01.amd.com> Co-authored-by: root <root@ctr-cx63-mi300x-21.amd.com> Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> Co-authored-by: root <root@gt-ccs-aus-h17-18.cs-aus.dcgpu> Co-authored-by: Cursor <cursoragent@cursor.com> |
||
|
Yung-sheng Tu
|
3ccb72e761 |
[rocm-libraries] ROCm/rocm-libraries#6207 (commit cc56378)
[CK TILE] Unification Work – Add `print()` Utility to `MmaOpTraits` (#6207) ## Motivation It would be useful to have a `print()` utility inside of unification work's code scope, so that we can print all template params and derived params of `amdgcn_mma` for easier debugging. ## Technical Details Adding helper functions and struct to traits, adding `print_flags()` for each `Default*CtrlFlags`, `amdgcn_target` and `MmaOpTraits` structs, and adding `print()` for `amdgcn_mma`. Note: the first commit is **not** in the scope of this PR. This PR should be merged after https://github.com/ROCm/rocm-libraries/pull/5801 and https://github.com/ROCm/rocm-libraries/pull/5857. ## Test Plan Adding test in layout test. ## Test Result Test should pass. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Vidyasagar Ananthan
|
b20458e19e |
[rocm-libraries] ROCm/rocm-libraries#5260 (commit a1834d2)
[CK] [CK_Tile] Add FMHA scaffolding to CK kernel dispatcher (#5260) ## Motivation The CK Tile dispatcher currently supports GEMM and Grouped Convolution but has no support for Fused Multi-Head Attention (FMHA). The example/ck_tile/01_fmha folder contains a comprehensive FMHA implementation with forward, backward, split-KV, paged-KV, append-KV, and batch-prefill kernels across multiple GPU architectures — but there is no unified dispatch layer for it. This PR ports the FMHA stack into the dispatcher, following the same architectural patterns established by GEMM and Grouped Convolution, enabling runtime kernel selection, JIT compilation from Python, and a declarative C++ example flow. Autotuning heuristics to follow. ## Technical Details This PR adds FMHA scaffolding to the CK dispatcher framework, mirroring GEMM's layered architecture. Seven new C++ runtime headers provide type definitions (coexisting with upstream headers via __has_include, requiring zero modifications to example/ck_tile/01_fmha/), a problem builder with 18+ setters, Signature + Algorithm kernel key matching, a virtual kernel instance, a DECL_FMHA_KERNEL_SET macro with wildcard support and named tile/wave/warp setters, arch-aware registry with JSON export, and a dispatcher with seqtune-aware selection, configurable timing, and multi-stage execution plans for split-KV (two-stage) and backward (three-stage). The codegen pipeline is driven by a fmha_arch_specs.json capturing per-arch tile tables and pipeline constraints for five architectures (gfx90a/942/950/1100/1201), migrated from hardcoded logic in 01_fmha/codegen/, with supporting modules for C++ symbol mappings, validation rules, and named receipt profiles (ck_default, flash, pytorch, aiter, fp32, fp8). Python integration (fmha_utils.py) mirrors the C++ layer with JIT compilation, parallel multi-kernel builds, HIP memory management via ctypes, tolerance-based validation, and a NumPy CPU reference with GQA support. Twenty-seven C++ and thirty-two Python examples cover the full feature surface — forward, split-KV, masks, bias, dropout, GQA, backward, append-KV, batch prefill, fp8, logits soft cap, sink tokens, and parameter sweeps — all JIT-compiled on the fly. ## Test Plan Seven test files cover the runtime types, codegen, and end-to-end correctness. C++ unit tests validate the problem builder, dispatcher planning (single-stage for forward/paged-KV/append-KV; multi-stage for split-KV and backward), registry operations, and the kernel-set declaration macro. Python unit tests verify codegen emission, profile filtering, and 15 validation rules for masks, hdim constraints, and pipeline requirements. GPU execution validation in 01_basic_fmha --validate reports zero errors across 65,536 elements with max absolute error of 7.29e-05. A gold-standard parity suite (test_fmha_parity.py) runs 14 configurations through both the upstream tile_example_fmha_fwd and the dispatcher, comparing exit codes to confirm behavioral parity — all 14 match. ## Test Result The C++ smoke test builds and passes all 9 compiled examples, and a Python JIT sweep (29_sweep_seqlen.py) passes 7/7 configurations reaching up to 375 TFLOPS at seqlen 2048. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Yaswanth Raparti <113389104+yraparti@users.noreply.github.com> Co-authored-by: Mohsen Saffari <mohsen.saffari@amd.com> Co-authored-by: Maksim (Max) Podkorytov <Maksim.Podkorytov@amd.com> Co-authored-by: yashagar <yashagar@amd.com> |
||
|
Bartłomiej Kocot
|
cc5c79a1e7 |
[rocm-libraries] ROCm/rocm-libraries#5904 (commit f4e261a)
[CK][CK Tile] Grouped Conv Backward Weight Streamk instances (#5904) ## Motivation Add streamk instance to grouped convolution backward weight profiler. ## Technical Details - New instances for grouped conv backward weight with streamk ## Test Plan test_grouped_convnd_bwd_weight_tile ## Test Result passed locally ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Graner, Johannes <johannes.graner@amd.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> |
||
|
Illia Silin
|
e496f94445 |
[rocm-libraries] ROCm/rocm-libraries#7516 (commit de93737)
[CK] fix some jenkins logic errors. (#7516) ## Motivation After merge from internal repo got some logic errors in the internal CI jenkinsfile. Here are 2 fixes for 2 issues: 1. make sure .ninja_log file exists before trying to parse it 2. use the default compiler for the gfx1250 target, since it's getting built in its own special docker, which does not have the option of installing alternative compilers. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Jobbins
|
8ca5c0a8d7 |
[rocm-libraries] ROCm/rocm-libraries#6961 (commit 47e8768)
[CK] print hostname and $NODE_NAME to find inconsistencies (#6961) ## Motivation We suspect that the check for amdgpu: `cat /sys/module/amdgpu/version` sometimes gets ran on the Jenkins controller instead of the node. This adds the `hostname` command to compare to the $NODE_NAME variable. ## Technical Details Updated Jenkinsfile to include the `hostname` command. ## Test Plan ## Test Result ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
John Shumway
|
ffebf1215b |
[rocm-libraries] ROCm/rocm-libraries#7114 (commit ecef372)
[CK] Add rocm_ck foundation types: DataType, Layout, Args, Ops (#7114) ## Summary - Add the vocabulary types that all rocm_ck schema headers build on - 9 new headers under `include/rocm_ck/`, 6 unit test files - Pure C++20, host-only — no CK Tile dependencies **Headers:** | Header | Purpose | |--------|---------| | `index_t.hpp` | `index_t`, `long_index_t` (matches ck_tile) | | `gpu_target.hpp` | `GpuTarget` enum (ISA targets) | | `datatype.hpp` | `DataType` enum (17 variants) | | `layout.hpp` | `Layout` enum (Row, Col, Auto) + stride helpers | | `fixed_string.hpp` | `FixedString<N>` — structural string for NTTPs | | `args.hpp` | Generic kernel argument buffer (ABI) | | `ops.hpp` | Operator structs (`GemmOp`, `AddOp`, ...) + `Op` variant | | `physical_tensor.hpp` | `PhysicalTensor` — maps names to Args slots | | `resolved_tensor.hpp` | `ResolvedTensor` — output of `Signature::resolve()` | **Stack**: This is PR 1 of 3 porting the rocm_ck constexpr schema from experimental to production, #7143. 1. **This PR** — Foundation types (vocabulary) 2. Schema engine — `Signature`, `resolve()`, `ArchProperties` 3. Spec factories — `GemmSpec`, `ElementwiseSpec`, `makeSpec()` ## Test plan - [ ] `ninja build-smoke-rocm-ck` builds all tests - [ ] `ctest -L ROCM_CK_SMOKE --output-on-failure` — 6 unit tests pass (86 test cases) - [ ] Default CK build (`CK_ENABLE_ROCM_CK=OFF`) unaffected 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Illia Silin
|
df07f060c1 |
[rocm-libraries] ROCm/rocm-libraries#7471 (commit 13b9eec)
[CK] increase timeout limit for fmha_fwd tests to avoid CI failure on gfx11 (#7471) ## Motivation This should prevent fmha_fwd tests from timing out on one of the slower gfx11 CI nodes and generating false CI failures. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
peterjunpark
|
da0ae0f97e |
[rocm-libraries] ROCm/rocm-libraries#7380 (commit 50b369d)
docs(composablekernel): update install pages for 7.13 (#7380) ## Motivation Update instructions to install prebuilt packages. Generally, point to ROCm Core SDK install instructions. List a more granular package if useful to users. Preview: https://rocm.docs.amd.com/projects/composable_kernel/en/users-peterjunpark-ck-7.13-install/install/Composable-Kernel-install.html >[!NOTE] >Some links appear as plain text in the preview b/c the target page isn't publicly accessible yet. Related to: - [x] composablekernel https://github.com/ROCm/rocm-libraries/pull/7380 - [x] hipblas https://github.com/ROCm/rocm-libraries/pull/7378 - [x] hipblaslt https://github.com/ROCm/rocm-libraries/pull/7379 - [x] hipcub https://github.com/ROCm/rocm-libraries/pull/7377 - [x] hipdnn https://github.com/ROCm/rocm-libraries/pull/7376 - [x] hipfft https://github.com/ROCm/rocm-libraries/pull/7375 - [x] hiprand https://github.com/ROCm/rocm-libraries/pull/7374 - [x] hipsolver https://github.com/ROCm/rocm-libraries/pull/7371 - [x] hipsparse https://github.com/ROCm/rocm-libraries/pull/7373 - [x] hipsparselt https://github.com/ROCm/rocm-libraries/pull/7372 - [x] miopen https://github.com/ROCm/rocm-libraries/pull/7370 - [x] rocblas https://github.com/ROCm/rocm-libraries/pull/7369 - [x] rocfft https://github.com/ROCm/rocm-libraries/pull/7368 - [x] rocprim https://github.com/ROCm/rocm-libraries/pull/7367 - [x] rocrand https://github.com/ROCm/rocm-libraries/pull/7366 - [x] rocsolver https://github.com/ROCm/rocm-libraries/pull/7364 - [x] rocsparse https://github.com/ROCm/rocm-libraries/pull/7365 - [x] rocthrust https://github.com/ROCm/rocm-libraries/pull/7363 - [x] rocwmma https://github.com/ROCm/rocm-libraries/pull/7362 <!-- Explain the purpose of this PR and the goals it aims to achieve. --> ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## JIRA ID <!-- If applicable, mention the JIRA ID resolved by this PR (Example: Resolves SWDEV-12345). --> <!-- Do not post any JIRA links here. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Bartłomiej Kocot
|
067e5e0ca4 |
[rocm-libraries] ROCm/rocm-libraries#6838 (commit ff7a665)
[CK_TILE] Add depthwise conv2d forward kernel (FP16/FP32) (#6838) ## Motivation CK currently has no kernel optimized for depthwise convolution (G=C_in=C_out, C=K=1 per group) and existing generic paths perform poorly for this workload. This PR adds a dedicated depthwise conv forward kernel in CK Tile. ## Technical Details Adds a dedicated depthwise conv2d forward op to CK Tile that performs direct convolution rather than falling back to the generic GEMM path. The kernel is templatized by filter size, stride, and data type, and compiled into ~60 instances covering common configurations (kernel 3/5/7/9, stride 1/2, FP16/FP32). Supports both CDNA (gfx942/gfx950) and RDNA (gfx1100/gfx1200) architectures. ## Test Plan - [x] Correctness and performance validated on gfx942, gfx950, and gfx1100, with ckProfiler `grouped_conv_fwd` as baseline. - [ ] MI300A (gfx942) and gfx1200 validation. ## Submission Checklist - [x ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. AICK-1137 --------- Co-authored-by: GenDu <Gen.Du@amd.com> |
||
|
Illia Silin
|
717f2efef7 |
[rocm-libraries] ROCm/rocm-libraries#6978 (commit e58096d)
[CK] add composable kernel support on gfx1250 (#6978) ## Motivation Add composable kernel support on gfx1250. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Qun Lin <qlin@amd.com> Co-authored-by: jialuo12_amdeng <jia.luo@amd.com> Co-authored-by: Andriy Roshchenko <andriy.roshchenko@amd.com> Co-authored-by: hsivasun_amdeng <haresh.sivasuntharampillai@amd.com> |
||
|
Illia Silin
|
ac18460782 |
[rocm-libraries] ROCm/rocm-libraries#7384 (commit 10e9d70)
[CK] Suppress new staging compiler errors (#7384) ## Motivation This should make new builds with staging compiler pass. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Yaswanth Raparti
|
caeeaf9cbd |
[rocm-libraries] ROCm/rocm-libraries#7289 (commit e3fb4ee)
[CK] Fix smart build false positives from merged commits (#7289) ## Motivation Current smart-build infrastructure triggers full build for almost every PR which is draining our CI infrastructure. Need to update the test selection logic based on diffs from the current workspace instead of entire repo. ## Technical Details Use three-dot syntax and scope BUILD_INFRA_PATTERN to composablekernel. Changes: - Switch from two-dot (..) to three-dot (...) in git diff - Three-dot shows only PR-specific changes - Excludes commits merged from develop (prevents false positives) - Scope BUILD_INFRA_PATTERN to projects/composablekernel/ paths only - Avoids triggering on other projects (hipblas, hipdnn, etc.) - Only composablekernel build infra changes trigger full build - Update both ci_safety_check.sh and validate_pr.sh ## Test Plan Test with PR 7112 and 7223 ## Test Result Impact: - PR 7112: Was 620 files (false positive) → Now 6 files (correct) - PR 7223: Was full build (false positive) → Now selective build (correct) ## Submission Checklist - [ x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
John Shumway
|
142ee00585 |
[rocm-libraries] ROCm/rocm-libraries#7090 (commit 316fded)
[CK] Add rocm_ck directory structure with feature flag (#7090) ## Summary Adds initial rocm_ck directory structure, #7119. - Establishes production `rocm_ck/` directory at `composablekernel/rocm_ck/`, peer to `tile_engine/` and `dispatcher/` - Adds `CK_ENABLE_ROCM_CK` option (default OFF) as a CK-internal feature flag — no superbuild or TheRock changes needed - Creates `rocm_ck` INTERFACE library, `ck_tile_headers` target, GTest integration with builder-style convenience targets (`smoke-rocm-ck`, `check-rocm-ck`) - Adds Jenkins `RUN_ROCM_CK_TESTS` parameter for CI, following the `RUN_BUILDER_TESTS` pattern - README explains the constexpr schema model: host-device separation via constexpr data rather than template parameters, enabling multi-arch distribution through kpack archives ## Test plan - [x] `cmake -DCK_ENABLE_ROCM_CK=ON` configures without errors - [x] `ninja check-rocm-ck` passes (4 host-only index type tests) - [x] Default build (`CK_ENABLE_ROCM_CK=OFF`) is unaffected — no rocm_ck targets present - [x] Jenkins `RUN_ROCM_CK_TESTS=true` enables the flag and runs `check-rocm-ck` 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com> |
||
|
Meekail Zain
|
0937b002d8 |
[rocm-libraries] ROCm/rocm-libraries#6867 (commit 3cb0219)
Added custom FMHA codegen receipt for TransformerEngine (#6867) ## Motivation TE uses AITER to build static MHA libraries, which ultimately rely on CK kernels. We use the `600` receipt which generates more kernels than TE truly needs. This bespoke receipt allows us to minimize the kernel count, compile time, and memory footprint of our MHA library. ## Technical Details Extended the receipt mechanism to include a custom `700` receipt for TE's needs ## Test Plan Test by building TE using the same receipt profile ## Test Result Build validated in TE using a custom feature branches of AITER/CK to temporarily apply the patch ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com> Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com> |
||
|
Yi DING
|
af7118e342 |
[rocm-libraries] ROCm/rocm-libraries#7331 (commit 5692db0)
[CK_TILE] Add async workspace prepare to FMHA BWD launcher (#7331) ## Motivation `aiter::mha_bwd` in group mode currently issues two synchronous `hipMemcpy` D2H copies to read `seqstart_q/k` for launcher construction. These sync copies block the host (~10–30 µs each) and implicitly synchronize the device by draining the stream, breaking CPU/GPU overlap on hot training paths. This PR adds a fully stream-async workspace preparation path on the FMHA BWD launcher so callers can pre-allocate the device workspace from upper-bound shapes and stage seqstart-dependent metadata via D2H/host-pack/H2D entirely on the user's stream. ## Technical Details - `FmhaBwdWorkspaceManager::GetWorkspaceDeviceSizeUpperBound` (`include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp`): computes the worst-case device dq_acc size from `(max_batch, hdim_q, nhead_q, max_seqlen_q, max_seqlen_k)` without dereferencing any seqstart array. Mirrors `PrepareWorkspaceHost`'s return value with worst-case bounds. - `fmha_bwd_launcher::prepare_workspace_async` (`example/ck_tile/01_fmha/fmha_bwd.hpp`): on the caller's stream, in order: 1. `hipMemsetAsync` of the dq_acc region (when `NeedsZeroDqAcc()`) 2. group mode: `hipMemcpyAsync` D2H of `seqstart_q/k` into a pinned host staging buffer 3. `hipLaunchHostFunc` runs `PrepareWorkspaceHost` on the pinned buffer 4. `hipMemcpyAsync` H2D of the packed metadata into `device_ws_ptr` The pinned staging buffer is held via `std::shared_ptr<void>` returned by a caller-provided `pinned_host_alloc` callback. Lifetime is extended past stream completion by a tail `hipLaunchHostFunc` scheduled in the launcher's destructor. - `ck_tile::pinned_host_releaser` (`include/ck_tile/host/pinned_host_releaser.hpp`): worker-thread utility for callers using bare `hipHostMalloc`. Defers `hipHostFree` off the HIP driver callback thread, which holds runtime locks and would deadlock against concurrent main-thread `hipFree`. PyTorch's `CachingHostAllocator` does not need this. - Example runner (`example/ck_tile/01_fmha/fmha_bwd_runner.hpp`): switched to the async path. ## Test Plan - `tile_example_fmha_bwd` (gfx950, dev preset `-Werror -Weverything`): - batch + nondet / batch + det / group + nondet / group + det - group + det 4-batch varlen (`-b=4 -h=8 -s=4096,3072,2048,1024 -d=128`) - FA (`flash-attention`) integration on ROCm 7.1.1 + PyTorch 2.9.1: - `tests/test_flash_attn_ck.py::test_flash_attn_varlen_deterministic` - `tests/test_flash_attn_ck.py::test_flash_attn_bwd_varlen_seqq_zero` ## Test Result - All CK runner cases `valid:y`. - FA pytest: **1952 passed in 44.82s**. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Copilot
|
c39bff93d0 |
[rocm-libraries] ROCm/rocm-libraries#6983 (commit f4e9a84)
Remove batch_prefill from FMHA_FWD_KNOWN_APIS (#6983) Remove `batch_prefill` from the `FMHA_FWD_KNOWN_APIS` list in `projects/composablekernel/example/ck_tile/01_fmha/CMakeLists.txt`. **Change:** ```cmake # Before set(FMHA_FWD_KNOWN_APIS "fwd;fwd_splitkv;fwd_appendkv;pagedkv_prefill;batch_prefill") # After set(FMHA_FWD_KNOWN_APIS "fwd;fwd_splitkv;fwd_appendkv;pagedkv_prefill") ``` Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: asleepzzz <4926646+asleepzzz@users.noreply.github.com> Co-authored-by: asleepzzz <hanwen.chang@amd.com> Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com> |
||
|
Qianfeng
|
0c63d6e776 |
[rocm-libraries] ROCm/rocm-libraries#7256 (commit 1fc20eb)
Skip numeric drop-out when PComputeWindow is a null_tile_window in Bl… (#7256) The BlockDropout implementation already provides very complete logic for generating random numbers and executing dropout for the P tensor after first attention Gemm with capability to support both Warp-Gemm 32x32 and 16x16 as well as to run on both wave32 and wave64 arch. But in some situation, we only need the block-layer process to generate random numbers, rather than simultaneously execute dropout in real-time on the vgpr tile. For example, xformers' `test_mem_eff_attention.py::test_dropout_ck` requires the host reference implementation of `attention forward with dropout` to use the same random numbers to compare & verify the device side implementation of `attention forward with dropout`, so a standalone kernel to generate random numbers only is required. This PR will enable xformers's random_val generating kernel (in file `ck_tiled_rand_uniform_kernel.h`) to depend on BlockDropout's `Run()` operator completely to generate random numbers for a `[MPerBlock, NPerBlock]` tile during the tile iteration, no need to replicate the logic of BlockDropout in the xformers kernel |
||
|
Linjun-AMD
|
5003f7ef8a |
[rocm-libraries] ROCm/rocm-libraries#7272 (commit d02f3c0)
[ck_tile][fmha_bwd] Fix sink_host OOB in group mode reference runner (#7272) ## Summary In `fmha_bwd_runner.hpp`, the `sink_host` `HostTensor` is allocated with first dimension `shape_batch` (= 1 in group mode), but the reference forward loop accesses `sink_host(wb, i_h)` with `wb ∈ [0, batch-1]`. For any `wb >= 1` this is an out-of-bounds heap read, silently corrupting the reference forward math chain (`lse_host`, `o_host`) and turning the bwd-side `d_sink_head_acc` reference into non-deterministic garbage. `HostTensor::operator()` does not bounds check, so the OOB is not caught at runtime. This manifests as intermittent `tile_example_fmha_bwd` failures (25–67% fail rate) when `-sink_grad=1` is combined with `-mode=1` (group mode), with bit-exact but spurious `max_err` values like 4.27 / 14.6. ## Fix One-line: allocate `sink_host` with `batch` (the real per-batch dim) instead of `shape_batch`, mirroring how `sink_host` is accessed by the loop. ```diff - sink_grad ? std::array<ck_tile::index_t, 2>{shape_batch, nhead} + sink_grad ? std::array<ck_tile::index_t, 2>{batch, nhead} Repro tile_example_fmha_bwd -b=2 -h=2 -s=516 -s_k=253 -prec=bf16 -d=72 \ -bias=n -dbias=0 -p_drop=0 -iperm=1 -operm=1 -deterministic=0 \ -v=3 -mode=1 -kname=1 -sink_grad=1 Verification - 0/30 fail on the repro config after fix - Baselines (before fix): - sink=1, mask=n: 25% fail rate (p ≈ 1.8e-4) - sink=1, mask=t: 67% fail rate (p ≈ 6e-15) Attribution Shape bug introduced together with sink_grad in #5504. Unrelated to #6914 (which is a fwd-only fix on a different code path) ``` ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Signed-off-by: junlin12 <junlin12@amd.com> Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com> |
||
|
Yaswanth Raparti
|
017dca1b9d |
[rocm-libraries] ROCm/rocm-libraries#6327 (commit 1e7a12e)
[CK][CK TILE] Dispatcher kernel selection heuristic for grouped conv (#6327) ## Motivation The ML heuristic in dispatcher does not support grouped-conv operator yet. In this PR, the support for fwd, bdw-data, and bwd-weight grouped-conv kernels have been added. A tile_engine utility has also been added to compile and run any selected kernel configuration through dispatcher infrastructure. ## Technical Details 1. Tile engine utility is added to benchmark each shape with all the possible kernel+tile_size combinations here - [https://github.com/ROCm/rocm-libraries/blob/users/yraparti/ck/dispatcher-grouped-conv-heuristics/projects/composablekernel/tile_engine/ops/grouped_conv/grouped_conv_full_benchmark.py](url) 2. New LGBM regressor models for grouped conv are added to models directory. We have 3 separate models for fwd, bwd-data, and bwd-weights [https://github.com/ROCm/rocm-libraries/tree/users/yraparti/ck/dispatcher-grouped-conv-heuristics/projects/composablekernel/dispatcher/heuristics/models](url) 3. Implemented lazy GPU initialization (dispatcher/python) - **Issue**: ProcessPoolExecutor fork() + GPU context caused memory access faults - **Solution**: Mirror FMHA pattern - defer GPU initialization until first run() - **Changes**: - setup_multiple_grouped_conv_dispatchers() returns List[Path], not loaded libs - GpuGroupedConvRunner.__init__() no longer calls ctypes.CDLL - Added _ensure_initialized() method for lazy GPU loading - GPU context created only on first run() call - **Benefit**: Parallel compilation now works without GPU conflicts 4. Addressed few miscellaneous issues such as: - Fixed BF16->FP16 naming bug in the dispatcher wrapper - Added new tile sizes, and comp_v5 pipeline to the arch spec to expand the kernel selection - Added automatic padding support for unsupported shapes in dispatcher runner - Created a single source of truth between tile_engine and dispatcher about the architecture and tile_size details - Build a validation scripts to compare oracle_best vs ml_heuristic comparison ## Test Plan 1. Validated fwd, bwd-data, and bwd-weight kernels with both known and unseen data sets with up to 300 problems. 2. Ensured that test cases are added in both dispatcher and tile_engine to validate the heuristic. ## Test Result Results on Unseen shapes validated on gfx950 #### Forward Pass Model - **Training Data**: 48,845 measurements across 1,372 unique problem shapes - **Validation Set**: 300 unseen problems from model crawler - **Validation Performance** (vs. oracle): - Mean Efficiency: **93.05%** - Median Efficiency: **96.8%** - P10 Efficiency: **79.9%** #### Backward Data Gradient (bwd_data) Model - **Training Data**: 18,773 measurements across 891 unique problem shapes - **Validation Set**: 300 unseen problems from model crawler - **Validation Performance** (vs. oracle): - Mean Efficiency: **93.8%** - Median Efficiency: **96.5%** - P10 Efficiency: **82.9%** #### Backward Weight Gradient (bwd_weight) Model - **Training Data**: 34,900 measurements across 1,508 unique problem shapes - **Validation Set**: 300 unseen problems from model crawler - **Validation Performance** (vs. oracle): - Mean Efficiency: **96.1%** - Median Efficiency: **99.2%** - P10 Efficiency: **89.4%** ## Submission Checklist - [ x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Vidyasagar Ananthan <vidyasagar.ananthan@amd.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Jan Patrick Lehr <JanPatrick.Lehr@amd.com> |
||
|
Illia Silin
|
22b9feb40f |
[rocm-libraries] ROCm/rocm-libraries#7111 (commit 651947f)
[CK] Fix latest batch of staging compiler warnings (#7111) ## Motivation Suppress the new batch of clang lifetimebound and invalidation warnings with the latest staging compiler. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Yi DING
|
370c7d762b |
[rocm-libraries] ROCm/rocm-libraries#7141 (commit 37e40c3)
[CK_TILE] Fix typo in fmha_fwd_kernel K-dram unmerge tuple sizes (#7141) ## Summary The qr_async_trload K-dram lambda's `else (XorLengthFold == 1)` branch in `fmha_fwd_kernel.hpp` writes the outer-tile dim of its 3-tuple unmerge/xor/merge as ```cpp number<FmhaPipeline::kQKHeaddim / kDramTileK / FmhaPipeline::kAlignmentK>{} ``` which divides one extra time. For every fp16/bf16 hdim=128 configuration the outer length collapses to **0**, e.g. `128 / 128 / 8 == 0`. The 3-tuple product no longer equals `kQKHeaddim`, so unmerge → xor → merge stops round-tripping the head dimension. This bug was masked by the async-load path: it only walks the descriptor via stride and silently absorbs a length=0 outer dim. Any consumer that actually traverses the descriptor (e.g. the TDM path on gfx1250) immediately faults on the resulting `tuple<int, constant<0>>`. The fix drops the extra `/ kAlignmentK` in all three call sites in the same lambda so the outer dim becomes `kQKHeaddim / kDramTileK` and the product is restored to `kQKHeaddim`. Strides are unaffected, so the async path is bit-identical. | Config (fp16/bf16) | hdim | kDramTileK | kAlignmentK | a (typo) | a (fixed) | product (typo) | product (fixed) | |---|---|---|---|---|---|---|---| | hdim128, kKLoadOnce | 128 | 128 | 8 | 0 | 1 | **0** | **128** | | hdim128, kK0=32 | 128 | 32 | 8 | 0 | 4 | **0** | **128** | | hdim64, kKLoadOnce | 64 | 64 | 8 | 0 | 1 | **0** | **64** | | hdim256, kK0=32 | 256 | 32 | 8 | 1 | 8 | **32** | **256** | Bug introduced in 2cc0af6a815a (PR #2888 \"[CK_TILE] FMHA FWD bug fix\"), where the original 2-tuple unmerge was generalized to a 3-tuple and the typo slipped in. ## Test plan - [x] Built `test_ck_tile_fmha_fwd` (umbrella, 5 gtest binaries) on gfx950 native at develop b3bdc63a509 with `dev-gfx950` preset (clang 22, ROCm 7.2.2). Compiles cleanly with `-Werror -Weverything`. - [x] Ran `ctest -R test_ck_tile_fmha_fwd` on gfx950 native, baseline vs patched: identical pass/fail (3 pass / 2 fail), identical failing case set (114 gtest fails + 2 GPU memory access faults, all in pre-existing fp16/bf16 group-mode `Alibi`/`Dropout` cases that reproduce on develop without this patch). Total wall time 403s → 393s. Per-case latency drift ±8% (noise). - [x] CI to verify on other gfx9 / gfx11 architectures. |
||
|
Illia Silin
|
b5d0cff36f |
[rocm-libraries] ROCm/rocm-libraries#6933 (commit ac8b7d9)
[CK] Filter out unsupported targets. (#6933) ## Motivation Filter out any unsupported targets, e.g., gfx900, gfx906, gfx90c, from the GPU_TARGETS or GPU_ARCHS lists. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Illia Silin
|
263e9965f6 |
[rocm-libraries] ROCm/rocm-libraries#7138 (commit 70e6660)
[CK] disable tile_engine by default, limit gfx1030 CI builds to develop only. (#7138) ## Motivation An attempt to reduce the build time and keep CI moving faster. Disable tile_engine by default since even the cmake step may take up to 30 minutes. Since we're down to a single gfx1030 CI node, use it only for develop builds. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Chao
|
3d8c21e838 |
[rocm-libraries] ROCm/rocm-libraries#6529 (commit 93a6097)
[CK_TILE] Enable V3 persistent kernel dispatch for FMHA forward on gfx950 (#6529) [CK_TILE] Enable V3 persistent kernel dispatch for FMHA forward on gfx950 ## Motivation Enable the existing V3 persistent kernel path for CK-Tile FMHA forward on gfx950 (MI350X/MI355X). The V3 kernel and codegen infrastructure already exist but are disabled via hardcoded `F_is_v3_enabled=False`. This change replaces the compile-time gate with a runtime environment variable `CK_FMHA_ENABLE_V3=1` (disabled by default, opt-in). When enabled: - **Prefill** workloads (seqlen_q > 1) dispatch to V3 persistent pipeline - **Decode** workloads (seqlen_q == 1) always use V2 (memory-bound, better suited) The V3 persistent kernel uses grid-stride scheduling, XCD-interleave tile assignment for L2 locality, LPT reversal for causal masks, and gfx950 async buffer loads. ## Technical Details Single file: `example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py` - Add `#include <cstdlib>` and `<string>` for `std::getenv` - Replace `{F_is_v3_enabled}` template parameter with runtime env var check - Add `seqlen_q > 1` guard (decode always uses V2) - Remove `.format()` call in `write_fwd_api()` ## Dependencies Depends on https://github.com/ROCm/rocm-libraries/pull/6501 — builds on XCD-interleave and LPT scheduling infrastructure. ## Test Plan - GPU validation on MI300X (gfx942, ROCm 6.4.1): - Command: `./build/bin/tile_example_fmha_fwd -b=2 -h=8 -s=4096 -d=128 -prec=bf16 -v=1 -warmup=1 -repeat=3` - GPU validation on MI350X (gfx950, ROCm 7.0): - Command (V2): `./build/bin/tile_example_fmha_fwd -b=2 -h=8 -s=4096 -d=128 -prec=bf16 -v=1 -warmup=1 -repeat=3` - Command (V3): `CK_FMHA_ENABLE_V3=1 ./build/bin/tile_example_fmha_fwd -b=2 -h=8 -s=4096 -d=128 -prec=bf16 -v=1 -warmup=1 -repeat=3` - Command (decode, always V2): `./build/bin/tile_example_fmha_fwd -b=64 -h=32 -h_k=8 -s=1 -s_k=4096 -d=128 -prec=bf16 -mode=group -v=1 -warmup=1 -repeat=3` ## Test Result Benchmark results (MI350X, gfx950, ROCm 7.0): | Config | V2 (TFlops) | V3 (TFlops) | Speedup | |--------|-------------|-------------|---------| | Non-causal b=2 h=8 hk=2 s=4096 d=128 bf16 | 696.3 | 884.2 | **+27.0%** | | Causal b=2 h=8 hk=2 s=4096 d=128 bf16 | 371.3 | 494.9 | **+33.3%** | | GQA b=2 h=32 hk=8 s=2048 d=128 bf16 | 671.3 | 831.7 | **+23.9%** | | LLaMA-70B b=1 h=64 hk=8 s=4096 d=128 bf16 | 761.5 | 927.3 | **+21.8%** | | Causal GQA b=2 h=32 hk=8 s=2048 d=128 bf16 | 345.4 | 631.9 | **+82.9%** | | Long-seq b=1 h=16 s=16384 d=128 bf16 | 797.8 | 969.9 | **+21.6%** | | Decode b=64 h=32 hk=8 s=1 s_k=4096 bf16 | 1828 GB/s | — (V2 path) | unaffected | Benchmark results (MI300X, gfx942, ROCm 6.4.1): V3 has 0% effect on MI300X — V3 relies on gfx950 async buffer loads and falls back to the V2 code path on gfx942. No regression on any config. | Config | TFlops / GB/s | Time (ms) | Delta vs baseline | |--------|-------------|-----------|-------------------| | MHA bf16 b=2 h=8 s=4096 d=128 | 342.98 TFlops | 0.401 | +0.1% | | MHA fp16 b=2 h=8 s=4096 d=128 | 411.18 TFlops | 0.334 | +4.9% | | Causal MHA bf16 b=2 h=8 s=4096 d=128 | 232.61 TFlops | 0.296 | +2.4% | | GQA 4:1 bf16 b=2 h=32 hk=8 s=2048 d=128 | 320.07 TFlops | 0.429 | -1.4% | | GQA 8:1 bf16 b=2 h=64 hk=8 s=2048 d=128 | 353.91 TFlops | 0.777 | +1.7% | | LLaMA-70B prefill b=1 h=64 hk=8 s=4096 d=128 bf16 | 381.53 TFlops | 1.441 | +1.2% | | Long-seq bf16 b=1 h=16 s=16384 d=128 | 388.61 TFlops | 5.659 | +1.4% | | Decode b=64 h=32 hk=8 s_k=4096 d=128 bf16 | 693.40 GB/s | 1.550 | +0.3% | All validation tests pass (`valid:y`) on both MI300X and MI350X. Additional validation: - `CK_FMHA_ENABLE_V3=0` correctly falls back to V2 (default behavior unchanged) - `CK_FMHA_ENABLE_V3=1` dispatches to V3 for prefill, V2 for decode - Validation passes across fp16/bf16, batch/group mode, causal/non-causal - No regression on decode path --------- Co-authored-by: Chao Zhou <chaozhou@fb.com> Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com> |
||
|
Linjun-AMD
|
cb61576896 |
[rocm-libraries] ROCm/rocm-libraries#6873 (commit b61b3fb)
[CK] add swiglustep_and_mul activation to gridwise_moe_gemm (#6873) Title： feat(composablekernel): add swiglustep_and_mul activation to gridwise_moe_gemm Description： ## Motivation Step-3.5-Flash uses a clamped SwiGLU activation (`swiglu_limits[43]=7`, `swiglu_limits[44]=7`) for layers 43 and 44. Without this kernel path, those layers produce BOS token spam because unclamped gate/up values accumulate floating-point noise over 200+ decode steps, degrading output quality (cosine similarity drops from 0.999989 to ~0.998982). ## Changes Add `swiglustep_and_mul` as a new `Activation` enum branch in `gridwise_moe_gemm.hpp`, covering all 4 code paths: - Quantized (A×B scale) + IsInputGemm=true - Quantized (A×B scale) + IsInputGemm=false - Non-quantized + IsInputGemm=true - Non-quantized + IsInputGemm=false The activation computes: gate = silu(gate) gate = clamp(gate, max=7.0f) up = clamp(up, min=-7.0f, max=7.0f) output = gate * up Also handles the `MulRoutedWeight` case (topk weight multiplication) and `pk_i4_t` weight scaling (×16 dequant factor). ## Verification - Tested on gfx950 (MI350X, 8×GPU) - cosine similarity for layers 43/44: **0.999989** (vs 0.998982 before fix) - End-to-end Step-3.5-Flash inference: no BOS spam, output coherent - BF16 tp=2/tp=4 and FP8 tp=2/tp=4 all verified PASS - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Linjun-AMD
|
5a2a362c46 |
[rocm-libraries] ROCm/rocm-libraries#6914 (commit b791478)
[CK_TILE][FMHA] Fix sink un-mask under right-window and emit fp8bf16 batch_prefill sink kernels (#6914) ## Summary Two related fixes to `ck_tile` FMHA so that StreamLLM-sink + sliding-window batch-prefill works correctly for fp8 KV / bf16 compute. Review the commits in this order: 1. `fmha: emit sink kernels for fp8bf16 batch_prefill` Extends `example/ck_tile/01_fmha/codegen/ops/fmha_batch_prefill.py` so the fp8(KV) / bf16(QO) batch-prefill codegen also emits the `mask=mask_enum::generic_with_sink` variant. Without this the runtime could not dispatch to a sink-aware kernel for the fp8bf16 path. 2. `fmha: respect right-window in IsOutOfSinkBound` The sink un-mask in `GenericAttentionMask::IsOutOfSinkBound` (local-mask branch) used `(i_y + x) > 1` as the gate, which conditioned on the row index instead of the column index. As a result, queries `1..sink-1` could attend to *future* sink positions (violating causal / right-window), while query `0` fell back to the plain causal mask. The fix replaces the guard with `i_x < i_y + x` so every query only sees sink columns up to its own right-window boundary. 3. `fmha: clarify IsOutOfSinkBound predicate comment` Doc-only follow-up that rewrites the comment above the predicate as a clause-by-clause explanation (`i_x < sink`, `i_x < i_y + x`, `y < y_total`, `i_y < x_total`). ## Test plan - [x] Repro on aiter `op_tests/test_batch_prefill.py` (fp8 + bf16_dequant modes with `sink=4`, `win_left=1023`, `softcap=0.0`, `sal=True`) now passes for all parametrized shapes. - [x] Existing fp16/bf16 batch-prefill paths (no sink) unchanged — codegen diff only adds the `generic_with_sink` variant for fp8bf16; existing kernel object lists unaffected. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: fengjunda.aml <fengjunda.aml@bytedance.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: root <root@smci350-rck-g03-f12-31.rck.dcgpu> |
||
|
Yi DING
|
2c677e8471 |
[rocm-libraries] ROCm/rocm-libraries#6152 (commit 36b016a)
[CK_TILE] Use Unified Workspace for FMHA BWD (#6152) ## Motivation `dq_acc` is the intermediate accumulation buffer used in FMHA backward pass for deterministic mode. The current implementation allocates it as a **single rectangular tensor**: ``` shape = [shape_batch, nhead, nsplits, shape_seqlen_q, hdim_q] ``` where `nsplits = launcher.dq_acc_splits` (a single scalar), computed from `max_seqlen_k` and shared across all batches. ### Problems 1. **Memory waste**: In group mode, each batch may have a different `seqlen_k`, but `nsplits` is computed from `max_seqlen_k`, causing batches with shorter `seqlen_k` to over-allocate in the split dimension. 2. **Interface coupling**: `fmha_bwd_args` exposes internal layout details such as `stride_dq_acc`, `nhead_stride_dq_acc`, `batch_stride_dq_acc`, and `split_stride_dq_acc`. The caller is responsible for computing these strides, but this logic belongs inside the kernel. ### Goals 1. Switch `dq_acc` buffer to a **compact layout**: batches are concatenated contiguously, with each batch occupying `nhead * nsplits_i * seqq_i * hdim_q` elements (nhead outermost). 2. **Remove all `*_stride_dq_acc` fields** from `fmha_bwd_args`, replacing them with a single `workspace_ptr`; the kernel splits this internally using a fixed layout. 4. `fmha_bwd_launcher` provides a **workspace management interface**: the caller only needs to allocate GPU memory and call `prepare_workspace()` — no layout computation required. 5. **Isolate kernel internals from the caller API**: the `dq_acc` layout (nsplits, strides, buffer size) is determined entirely inside the launcher/kernel. Future changes to block shape, pipeline type, or persistent kernel strategy require no modifications to the caller's `fmha_bwd_args` or workspace allocation logic. ## Technical Details ### Interface Design #### New fields in `fmha_bwd_traits` ```cpp struct fmha_bwd_traits { int seqlen_q; int seqlen_k; int batch; int max_seqlen_q; int max_seqlen_k; int hdim_q; int hdim_v; int nhead_q; int nhead_k; std::string data_type; bool is_group_mode; mask_enum mask_type; bias_enum bias_type; bool has_dbias; bool has_dropout; bool is_store_randval; bool is_deterministic; // New: cumulative physical seqlen pointers for group mode (pass nullptr for batch mode). // seqstart_qs[i+1] - seqstart_qs[i] = physical seqlen_q of batch i (including padding); length = batch+1 // seqstart_ks[i+1] - seqstart_ks[i] = physical seqlen_k of batch i (including padding); length = batch+1 const int* seqstart_qs = nullptr; const int* seqstart_ks = nullptr; }; ``` #### `fmha_bwd_launcher` actual structure ```cpp struct fmha_bwd_launcher { std::function<float(fmha_bwd_args, const ck_tile::stream_config&)> run{}; // Total workspace size in bytes (host_ws_size + device_ws_size), computed by init(). // Zero for kUseQrQtrDorPipeline (writes dq directly, no acc buffer needed). size_t workspace_size = 0; fmha_bwd_launcher(const fmha_bwd_traits&); // Copies auxiliary data (nsplits[], offsets[]) via hipMemcpy to the head of the GPU workspace, // and zeros the dq_acc buffer portion (tail of workspace) if required. // The memory pointed to by device_ws must be >= workspace_size bytes. std::function<void(void* device_ws)> prepare_workspace{}; template <typename... Args> float operator()(Args&&... args) const { return run(std::forward<Args>(args)...); } private: size_t host_ws_size = 0; // CPU workspace size (nsplits[] + offsets[] arrays) size_t device_ws_size = 0; // GPU-only data size (dq_acc buffer) std::unique_ptr<char[]> ws_host; // host-side workspace buffer public: template <typename T0, typename T1, typename T2, typename Arch> void init(const fmha_bwd_traits& traits); }; ``` The `init<>()` template method (invoked by codegen dispatch branches as `this->init<...>(t)`) is responsible for: 1. Setting the `run` lambda 2. Calling `FmhaBwdDQDKDVKernel::GetWorkspaceHostSize(batch)` to obtain `host_ws_size` 3. Allocating `ws_host` (host memory) 4. Calling `FmhaBwdDQDKDVKernel::PrepareWorkspaceHost(ws_host.get(), ...)` to fill nsplits/offsets; return value is `device_ws_size` 5. `workspace_size = host_ws_size + device_ws_size` 6. Setting the `prepare_workspace` lambda (captures `this`, calls `PrepareWorkspaceDevice`) When no kernel matches the given traits, both `run` and `prepare_workspace` are initialized to default lambdas that print a warning to `std::cerr` and return gracefully (no exception). #### Workspace overall layout The workspace is managed by `FmhaBwdWorkspaceManager` and consists of two segments: ``` Offset 0 (CPU-prepared segment, host_ws_size bytes; also hipMemcpy'd to the head of GPU workspace): index_t nsplits[batch or 1] — per-batch nsplits array group mode: batch elements batch mode / non-deterministic: 1 element [group mode only] long_index_t dq_acc_offsets[batch+1] — per-batch element offset (inclusive prefix sum) offsets[0]=0, offsets[i+1] = offsets[i] + nhead*nsplits_i*seqq_i*hdim_q Offset host_ws_size (device data segment, device_ws_size bytes): AccDataType dq_acc[total_elements] — compact dq_acc buffer (zeroed if required) total_elements = sum_i(nhead * nsplits_i * seqq_i * hdim_q) layout within each batch: [nhead, nsplits_i, seqq_i, hdim_q] note: seqq_i uses the physical length (including padding) ``` Alignment constant (`ALIGNMENT = 16`): ``` nsplits_size = align_up(sizeof(index_t) * N, 16) // N = batch (group) or 1 (batch/non-det) offsets_size = align_up(sizeof(long_index_t) * (batch+1), 16) // group mode only host_ws_size = nsplits_size + offsets_size dq_acc_offset = host_ws_size // GetDqAccDataOffset(batch) ``` **Key benefits**: - The kernel reads nsplits/offsets directly from the workspace head — no device-side recomputation. - `FmhaBwdConvertQGradKernel` is completely decoupled from the pipeline block shape (`kN0`): nsplits is read from `nsplits_ptr`, `kN0` is no longer a template parameter, and multiple dq_dk_dv tiles with different `F_bn0` values now share a single convert_dq kernel instance (under receipt 1/2, deterministic convert_dq kernel count drops from ~300 to 60). - nsplits/offsets are computed on the host and transferred in one `hipMemcpy`; the dq_acc buffer follows immediately, at the offset given by `GetDqAccDataOffset`. #### Workspace size by scenario | Scenario | `workspace_size` | Notes | |----------|-----------------|-------| | **kUseQrQtrDorPipeline** (any mode) | `0` | Writes dq directly; no acc buffer; `PrepareWorkspaceHost` returns 0 | | **Non-deterministic + batch mode** | `> 0` | nsplits[1]=1; dq_acc used for atomic add; `workspace_size = host_ws_size + batch*nhead*seqlen_q*hdim_q*ebytes` | | **Non-deterministic + group mode** | `> 0` | nsplits[1]=1; dq_acc contiguous layout; `workspace_size = host_ws_size + nhead*seqstart_qs[batch]*hdim_q*ebytes` | | **Deterministic + group mode** | `> 0` | nsplits[batch], offsets[batch+1], compact dq_acc; nsplits_i computed independently per batch | | **Deterministic + batch mode persistent** | `> 0` | nsplits[1] (uniform across batches); dq_acc `batch*nhead*nsplits*seqlen_q*hdim_q` | **NeedsZeroDqAcc** (determines whether `PrepareWorkspaceDevice` calls `hipMemset`): - Persistent kernel (deterministic batch mode) or non-deterministic: **must zero** (atomic add requires zero initialization) - Deterministic group mode + no mask: **no zeroing needed** (every tile writes its full region) - Deterministic + with mask: **must zero** (some blocks are skipped, leaving uninitialized tiles that would contribute to the reduction) #### Caller usage ```cpp // 1. Create launcher (traits include seqstart_qs/ks pointers; workspace_size is computed during construction) fmha_bwd_launcher launcher(fmha_traits); // 2. Read launcher.workspace_size directly const auto ws_size = launcher.workspace_size; // 3. Allocate a single GPU workspace ck_tile::DeviceMem ws_buf(ws_size); // 4. Copy nsplits/offsets to GPU head and zero dq_acc if required launcher.prepare_workspace(ws_buf.GetDeviceBuffer()); // 5. Build args with a single workspace pointer; the kernel splits it internally fmha_bwd_args args{ ..., ws_size > 0 ? ws_buf.GetDeviceBuffer() : nullptr, // workspace_ptr }; launcher(args, stream_config); ``` --- ### Key Code Structure #### FmhaBwdWorkspaceManager (`fmha_bwd_kernel.hpp`, new class) ```cpp template <typename AccDataType, bool kIsGroupMode, bool kIsDeterministic> struct FmhaBwdWorkspaceManager { static constexpr size_t ALIGNMENT = 16; // CPU workspace (nsplits + offsets) sizes static size_t GetDqAccSplitsSize(int batch); // align_up(sizeof(index_t)*N, 16) static size_t GetDqAccOffsetsSize(int batch); // group mode only: align_up(sizeof(long_index_t)*(batch+1), 16) static size_t GetWorkspaceHostSize(int batch); // = SplitsSize + OffsetsSize // Starting offset of dq_acc data within the full workspace (= host_ws_size) static size_t GetDqAccDataOffset(int batch); // = GetWorkspaceHostSize(batch) // Fills nsplits/offsets in the CPU workspace; returns device_ws_size (dq_acc buffer bytes) template <bool kUseQrQtrDorPipeline, index_t kN0> static size_t PrepareWorkspaceHost(void* cpu_ws, index_t batch_size, index_t hdim_q, index_t nhead_q, index_t seqlen_q, index_t seqlen_k, const index_t* seqstart_qs, const index_t* seqstart_ks); // hipMemcpy's cpu_ws to device_ws head; hipMemset's the dq_acc portion to 0 if required template <bool kUseQrQtrDorPipeline, bool kHasMask> static void PrepareWorkspaceDevice(void* device_ws, const void* host_ws, size_t device_ws_size, size_t host_ws_size); }; ``` #### workspace_ptr parsing (inside the kernel) The kernel parses three address regions from `kargs.workspace_ptr`: **Group mode (`FmhaBwdDQDKDVKernel::MakeKargs`)**: ```cpp const uint8_t* ws = reinterpret_cast<uint8_t*>(workspace_ptr); // dq_acc_ptr (stored in FmhaBwdCommonKargs) ws + WorkspaceManager::GetDqAccDataOffset(batch) // dq_acc_batch_offset_ptr (FmhaBwdGroupModeKargs field) reinterpret_cast<const long_index_t*>(ws + WorkspaceManager::GetDqAccOffsetsOffset(batch)) ``` **Batch mode**: ```cpp ws + WorkspaceManager::GetDqAccDataOffset(batch) // dq_acc_ptr // No offsets pointer; batch offset is computed inside run_() from nsplits ``` **`FmhaBwdConvertQGradKernel`** follows the same pattern: - Group mode: extracts `dq_acc_ptr`, `dq_acc_batch_offset_ptr`, and `nsplits_ptr` (`GetDqAccSplitsOffset(batch)`) from workspace - Batch mode: reads nsplits from `nsplits_ptr[0]`; batch offset computed internally ### Addressing in `run_()` (group mode) ```cpp // Per-batch processing: const long_index_t batch_offset_dq_acc = kargs.dq_acc_batch_offset_ptr[i_batch]; // seqq_i (physical length) derived from seqstart_q_ptr const index_t seqq_i = kargs.seqstart_q_ptr[i_batch+1] - kargs.seqstart_q_ptr[i_batch]; // nsplits_i read from nsplits_ptr (convert_dq kernel) or from GetDqAccSplits const long_index_t split_stride_i = static_cast<long_index_t>(seqq_i) * kargs.hdim_q; const long_index_t nhead_stride_i = static_cast<long_index_t>(nsplits_i) * split_stride_i; // Final address: dq_acc_base + batch_offset_dq_acc + i_nhead * nhead_stride_i + i_split * split_stride_i ``` #### nsplits computation (`PrepareWorkspaceHost`) `PrepareWorkspaceHost` is a template method of `FmhaBwdWorkspaceManager` that still takes `kN0` as a template parameter (from `BlockFmhaShape::kN0` of the dq_dk_dv pipeline). However, this parameter is **only used inside this host-side function** to compute nsplits — it is no longer passed into the convert_dq kernel. | Mode | nsplits computation | |------|---------------------| | kUseQrQtrDorPipeline | Writes dq directly; nsplits[0]=0; returns device_ws_size=0 | | Non-deterministic | nsplits[0]=1; dq_acc used for atomic add | | Deterministic + group mode | `ceil((seqstart_ks[i+1]-seqstart_ks[i]) / kN0)` computed per batch | | Deterministic + batch mode persistent | Same logic as the original `GetDqAccSplits` (`dqdqkdv_workers` based) | ### Removing kN0 dependency from `FmhaBwdConvertQGradKernel` `FmhaBwdConvertQGradKernel` previously required `kN0` as a template parameter (via `BlockFmhaBwdConvertQGradPipelineProblem`) for two purposes: 1. In batch mode `operator()`: self-computing `nsplits = ceil(seqlen_k / kN0)` 2. The `b{kM0}x{kN0}` component of the kernel name string Both have been removed in this refactor: - **Batch mode**: now reads `kargs.nsplits_ptr[0]` directly (guarded by `if constexpr(kIsDeterministic)` to avoid accessing a non-existent field in non-deterministic instances) - **Kernel name**: simplified to `b{kM0}`, no longer includes `kN0` - **Template parameters**: `BlockFmhaBwdConvertQGradPipelineProblem` drops the `kN0_` parameter; `fmha_bwd_convert_dq_traits_` drops the `kN0` parameter; `F_bn0`/`convert_dq_bn0` fields removed from codegen Effect: all dq_dk_dv tiles sharing the same `(hdim, dtype, mode, pad, deterministic)` combination — regardless of `F_bn0` value (16/64/128/192/256) — now share a **single** convert_dq kernel instance. --- ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Illia Silin
|
1f8dbfb63e |
[rocm-libraries] ROCm/rocm-libraries#7046 (commit aaf7665)
[CK] fix CI git token. (#7046) ## Motivation Fix the CI breakage due to git PAT deprecation. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Jeff Huang
|
faa9dc52cb |
[rocm-libraries] ROCm/rocm-libraries#6932 (commit ce3e67b)
[CK] Fix OOB page table read in batch_prefill V prefetch (AICK-1171) (#6932) ## Summary Fix a GPU memory access fault in `mha_batch_prefill` triggered when the per-batch page table is tightly sized (no trailing slack). **Affected configurations:** - All FMHA batch prefill V2 kernels (`block_fmha_batch_prefill_pipeline_qr_ks_vs_async`) - Triggered by paged KV layouts where `kv_page_indices.numel() == ceil(seqlen_k / page_size)` exactly - Manifests as: `Memory access fault by GPU node-X (Agent handle: 0x...)` followed by `Aborted (core dumped)` - Silent corruption (no fault, wrong output) when the OOB read happens to land in zero-initialized memory ### Root cause `load_physical_pages` performs **lookahead reads** on the page table to prefetch K/V tiles for the next iteration. When the page table for a batch has exactly `N` entries, the V-tile prefetch indexes `page_idx[N]` (one past the last valid entry), reading either uninitialized memory or the next batch's slot. On gfx942 with a tightly-sized page table, the read crosses into an unmapped page and triggers an HSA page fault. The bug was masked in earlier testing because most test harnesses pad `kv_page_indices` with trailing zeros — OOB reads then return `page_id = 0`, a valid in-cache page, producing silent numerical drift instead of a fault. ### Fix design Thread `max_page_table_idx = (seqlen_k - 1) / page_size` from the kernel layer down to `load_physical_pages`, and clamp every page-table read with `ck_tile::min()`. Applied to **all four code paths** in the V prefetch: | Branch | What it does | Clamp applied | |--------|-------------|---------------| | `kIsKcache` | K prefetch loop | `min(global_token_idx >> kLog2PageSize, max_page_table_idx)` | | V LINEAR (`page_size == 1`) | One token = one page | `min(global_token_idx, max_page_table_idx)` | | V crosses pages (`kVTileCrossesPages`) | Per-thread page lookup | `min(global_token_idx >> kLog2PageSize, max_page_table_idx)` | | V single page (lane0 broadcast) | `readfirstlane`-uniform lookup | `min(... >> kLog2PageSize, max_page_table_idx)` | ### Key design decisions **Mandatory parameter, not optional with a sentinel default.** An optional `max_page_table_idx = INT32_MAX` default would let the bug silently come back at any new callsite that forgets to pass it. Making it mandatory forces every caller to opt in explicitly and surfaces missed callsites at compile time. **`seqlen_k == 0` clamps to 0** instead of underflowing `(0 - 1) / page_size` to `-1`. The empty-batch case is rare but well-defined: clamp every read to slot 0. **Single computation in the kernel layer.** `FmhaBatchPrefillWithPagedKVCacheKernel` computes `max_page_table_idx` once per batch and forwards it through every QScale branch (PERTENSOR / KV_BLOCKSCALE / default). All three `operator()` overloads of the pipeline (rich, default forwarder, KV_BLOCKSCALE forwarder) take and forward the parameter. ### Files changed | File | Change | |------|--------| | `include/ck_tile/ops/fmha/kernel/fmha_batch_prefill_kernel.hpp` | Compute `max_page_table_idx` per batch, forward to all 3 QScale branches | | `include/ck_tile/ops/fmha/pipeline/block_fmha_batch_prefill_pipeline_qr_ks_vs_async.hpp` | Add `max_page_table_idx` to `load_physical_pages` and 3 `operator()` overloads; clamp page-id reads in 4 code paths | ## Test plan - [x] AICK-1171 reproducer verified on MI-308X (gfx942) - [x] New pytest case `test_batch_prefill_aick1171_oob_page_table_read` in aiter, parametrized over `total_blocks ∈ {160, 164, 168, 176, 208, 256}` (matches the `crash1_r8_*` bisect family) - [x] Full FMHA batch prefill suite on gfx942 + gfx950 ## Linked issue AICK-1171. |
||
|
Yaswanth Raparti
|
3cc95874f9 |
[rocm-libraries] ROCm/rocm-libraries#6912 (commit c705da2)
[CK] Reduce per-file logging in cmake_dependency_analyzer (#6912) ## Motivation Current progress_callback function generates large volume of prints which creates noise in seeing actual CI failure logs. Only emit a progress line at the completion of each stage to avoid massive logs from the per-source-file extracting_dependencies callback. ## Technical Details Update the `progress` function to print only at the completion of each stage. https://github.com/ROCm/rocm-libraries/pull/6912/changes#diff-15971b83c7dfefb48fd788507a923017d93bbd9487ed6aeb414ad2c5e00be934R720 ## Test Plan to be tested in CI ## Test Result to be tested in CI ## Submission Checklist - [x ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
John Shumway
|
0ec64c246c |
[rocm-libraries] ROCm/rocm-libraries#6972 (commit 8761b90)
[CK] Dockerfile: auto-discover latest TheRock nightly tarball (#6972) ## Motivation Our docker containers with `--build-arg compiler_version=therock` should have the latest nightly build of TheRock in `/opt/rocm`. When I looked for `rocm_kpack` and other `kpack` artifacts, they were missing, and I realized we had pinned the version by date. Instead, we should look for the most recent linux-multiarch tarball. ## Summary - Auto-discover the latest TheRock nightly tarball at Docker build time instead of pinning a stale URL (previously hardcoded to a Feb 2026 nightly that predates kpack) - Logic is to `wget` the directory, and identify the latest tarball (alphabetically sorted by YYYYMMDD in filename). - Support manual override via `--build-arg TARBALL_URL=...` for pinning, and `--build-arg TARBALL_PATTERN=...` for selecting a specific arch variant - Fix sccache download URL: `/releases/latest/download/` was redirecting to v0.15.0 but the filename referenced v0.14.0, causing a 404 ## Test plan - [x] Verified tarball discovery logic resolves to `therock-dist-linux-multiarch-7.13.0a20260430.tar.gz` - [x] Built Docker image locally with `--build-arg compiler_version=therock` - [x] Confirmed sccache installs successfully with the fixed URL - [ ] Verify CI pipeline builds with the updated Dockerfile 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
ltqin
|
501e7ef12a |
[rocm-libraries] ROCm/rocm-libraries#6574 (commit b3db057)
[CK_TILE] Add SageAttention v2 forward kernel with multi-granularity quantization (#6574) ## Summary Add a CK_TILE forward kernel implementing [SageAttention v2](https://arxiv.org/abs/2411.10958) — an attention algorithm that applies multi-granularity quantization to Q/K/V before computing attention, trading minimal accuracy loss for higher throughput on low-precision hardware. ### Quantization design | Tensor | Supported data types | Scale granularity options | |--------|---------------------|--------------------------| | Q | fp8 / int8 / int4 | per-tensor, per-block (128 tokens), per-warp (32 tokens), per-thread (4 tokens) | | K | fp8 / int8 / int4 | per-tensor, per-block (128 tokens), per-warp (64 tokens), per-thread (16 tokens) | | V | fp8 | per-channel (always) | | O | bf16 | — | Three precision combinations are supported: `fp8/bf16` (QKV fp8, O bf16), `i8/fp8/bf16` (QK int8, V fp8, O bf16), and `i4/fp8/bf16` (QK int4, V fp8, O bf16). ### Architecture support - **gfx9** (CDNA2/3, e.g. gfx90a, gfx942) — full tile set - **gfx950** (CDNA4) — restricted tile set (N-per-block capped at 64 for fp8-family dtypes) ### Implementation - Two pipeline variants: `QRKSVS` (synchronous) and `QRKSVS_ASYNC` (async copy) - Masking support: no mask, causal (top-left / bottom-right), and generic windowed - Batch and group (variable-length) modes - Head dimension: d=128, d_v=128 - Python codegen under `example/ck_tile/49_sageattention/codegen/` generates kernel instances per target/dtype/tile combination - Smoke tests included via `tile_example_sageattn_fwd` ### Test commands \`\`\`bash # fp8 QKV ./build/bin/tile_example_sageattn_fwd -v=1 -b=16 -h=8 -s=1024 -d=128 -kname=1 -prec=fp8bf16 -qscale=3 -init=3 # int8 QK, fp8 V ./build/bin/tile_example_sageattn_fwd -v=1 -b=16 -h=8 -s=1024 -d=128 -kname=1 -prec=i8fp8bf16 -qscale=3 -init=3 \`\`\` \`-qscale\` values: 1=per-tensor, 2=per-block, 3=per-warp, 4=per-thread |
||
|
ArthurLiu
|
67aa854621 |
[rocm-libraries] ROCm/rocm-libraries#6764 (commit 8c20d70)
[CK][CK_TILE] Fix FMHA codegen group mode dispatch (#6764) ## Motivation FMHA codegen had incorrect dispatch behavior in group mode. Two root causes: 1. Wrong field names in dispatch conditions — Used batch-mode fields (seqlen_q, seqlen_k) instead of group-mode fields (max_seqlen_q, max_seqlen_k), causing wrong kernel selection at runtime on gfx950. 2. Missing kernel variants — Group mode was overly filtered out from smaller-tile specializations (bwd) and lacked spatial-padding pipeline variants on gfx950 (fwd). gfx942 don't support trload pipeline. ## Technical Details fmha_bwd.py: - max_seq_q_cond and extra_cond now emit t.max_seqlen_q / t.max_seqlen_k for group mode. - Relaxed kernel filtering: group mode no longer skips tiles with max_seq_q != 0. fmha_fwd.py: - get_bm0_cond emits a.max_seqlen_q for group mode tile-size dispatch. - Added two qr_async_trload pipeline variants with spatial padding for gfx950 group mode. ## Test Plan Triggering AITER CI job: ## Submission Checklist - [ x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |
||
|
Wojciech Laskowski
|
640bd560ec |
[rocm-libraries] ROCm/rocm-libraries#5801 (commit 27f6d15)
[CK Tile] Adding WMMA wrappers for dense builtins (#5801) ## Motivation This PR is part of the [WMMA/MFMA] unification work. It's the first of the series of PRs that add all the necessary MMA builtins as a `amdgcn_mma` structs. ## Technical Details This change adds new specializations for WMMA dense builtins. In total, we have now 9 RDNA4 builtins and 3 RDNA3 builtins. ## Test Plan All the new wrappers were added to the test suite in `test_amdgcn_mma_layout.inc`. ## Test Result Test pass locally, waiting for the CI. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. --------- Co-authored-by: Yung-sheng Tu <yung-sheng@streamhpc.com> |
||
|
Illia Silin
|
5df88bfe73 |
[rocm-libraries] ROCm/rocm-libraries#6741 (commit 0d4180f)
[CK] restore fmha performance reporting and disable c++17 in CI. (#6741) ## Motivation This change restores monitoring of FMHA benchmarks performance in daily builds and removes the std=c++17 flag from CI builds on gfx90a. ## Technical Details <!-- Explain the changes along with any relevant GitHub links. --> ## Test Plan <!-- Explain any relevant testing done to verify this PR. --> ## Test Result <!-- Briefly summarize test outcomes. --> ## Submission Checklist - [ ] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. |