Fixing two cascading bugs when running the CK MoE tuner by xaguilar-amd · Pull Request #2464 · ROCm/aiter

xaguilar-amd · 2026-03-25T13:40:11Z

Problem

Running the CK MoE 2-stage tuner fails immediately with two cascading errors,
leaving zero shapes tuned:

[aiter] [Failed] Task 4 failed with OSError:
  aiter/jit/module_moe_asm.so: undefined symbol: _ZTVN5torch8autograd12AutogradMetaE

[aiter] error in batch 1 of 1: list index out of range
Traceback (most recent call last):
  File "aiter/utility/base_tuner.py", line 468, in run
    all_results = self.tune(batch, self.tunedf, args)
  File "csrc/ck_gemm_moe_2stages_codegen/gemm_moe_tune.py", line 2311, in tune
    rets = mp_tuner(...)
  File "aiter/utility/mp_tuner.py", line 535, in mp_tuner
    result.append(task_result[0])
IndexError: list index out of range

Root causes

1 — `module_moe_asm.so` was compiled without torch linkage (`aiter/jit/core.py`)

module_moe_asm is used in two modes: as a pybind extension (loaded via
importlib) and as a raw ctypes library (loaded via ctypes.CDLL).
The ctypes path lives in _ctypes_call._ensure_loaded() and, before this fix,
unconditionally forced:

d_args["torch_exclude"] = True

torch_exclude=True drops all torch link flags (-ltorch_cpu, -lc10, etc.)
from the build. This is intentional for purely standalone C kernel modules —
but module_moe_asm's source list includes csrc/pybind/moe_op_pybind.cu, a
pybind11 file that uses at::Tensor and other PyTorch C++ types. Those types
reference the vtable symbol _ZTVN5torch8autograd12AutogradMetaE, which the
linker left undefined:

$ nm -D aiter/jit/module_moe_asm.so | grep AutogradMeta
                 U _ZTVN5torch8autograd12AutogradMetaE

Because libtorch_cpu.so was absent from NEEDED, ctypes.CDLL failed when
loading the .so in tuner worker subprocesses. The tuner uses
mp.set_start_method("spawn"), so each worker starts a fresh Python
interpreter. Even though import torch runs inside the worker (via
mp_tuner.py's top-level import), PyTorch's own shared libraries are loaded
as transitive dependencies of torch._C.so. On Linux, transitive
dependencies do not inherit RTLD_GLOBAL, so their symbols are invisible to
subsequent ctypes.CDLL calls.

Fix: remove the forced override so each module is built with its own
configured value (which defaults to torch_exclude=False for
module_moe_asm). The rebuilt .so lists libtorch_cpu.so in NEEDED, and
the dynamic linker resolves all symbols automatically.

2 — Failed tasks left `result_dict` empty, causing `IndexError` (`aiter/utility/mp_tuner.py`)

mp_tuner collects worker results into result_dict[k] and reconstructs a
flat list at the end:

for k in range(len(rets)):
    task_result = result_dict.get(k, [])
    ...
    result.append(task_result[0])   # IndexError if task_result is []

When a task raises any exception that is not MPTimeoutError and not
KeyError, the existing handler logged the error and marked the task as
complete — but never wrote a placeholder into result_dict[k]. The loop
then called task_result[0] on an empty list and raised IndexError, crashing
the tuner before it could print which shapes failed.

Fix: add the same add_dummy_result call that already existed for the
timeout branch so every failed task gets a (info, float("inf"), 1.0)
placeholder in result_dict.

How the two bugs interact

Bug 1 caused every ASM-stage-1 task to raise OSError in the subprocess.
Bug 2 meant that the parent process then crashed with IndexError when
collecting results — before any summary or diagnostic output could be printed.

Fixing Bug 2 alone allows the tuner to survive and produce a proper
"Failed shapes" summary, but no ASM-stage-1 kernels are tuned. Both fixes are
required for correct operation.

Testing

After deleting the stale .so (so it is rebuilt with correct flags on the next
run) and applying both patches, the tuner completes without errors.

Notes for reviewers

The torch_exclude=True override in _ctypes_call._ensure_loaded is safe to
remove because any module that has torch_exclude explicitly set to True in
optCompilerConfig.json (e.g. module_aiter_enum) will continue to use that
value. Only modules that rely on the default (False) are affected.
If a future module genuinely needs both ctypes loading and a torch-free
build, its pybind source file should be separated from its pure-C sources into
a distinct build target so the two concerns do not conflict at link time.

PR created with Cursor.

github-actions · 2026-03-25T13:40:39Z

🏷️ CI Guide

Runs automatically on every PR:

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

Extended tests (opt-in via labels):

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

Add labels via the sidebar or gh pr edit 2464 --add-label <label>

@xaguilar-amd

@xaguilar-amd asked to drop #2464 (CK MoE tuner bug fixes) from this bulk merge — they don't need it for the uplift. Verified that #2464 is the only PR in this bulk merge touching aiter/jit/core.py and aiter/utility/mp_tuner.py: the diff between the branch and origin/main on those files is exactly #2464's +9/-1 and +5/-0, with no other PR content mixed in. Restoring both files to origin/main therefore drops #2464 cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@xaguilar-amd

…3-Next, pa_mqa OOB) (#3005) * fix: remap QuantType.No to per_1x32 for fp4x2 MoE weights (W4A6 support) * Fixing two cascading bugs when running the MoE tuner * Enable split-K for block-scale A8W8 CK and CKTile GEMMs Propagate the splitK parameter (as KBatch = 2^splitK) through the block-scale GEMM kernel infrastructure so that the tuning scripts can sweep split-K values to improve occupancy on small-M shapes. CK path: add KBatch parameter to gemm_a8w8_blockscale_impl and call SetKBatch on the device argument. The CK invoker handles output zeroing and atomic accumulation internally. CKTile path: add k_batch parameter to gemm_a8w8_blockscale_cktile_impl, remove the "split-k is not supported yet" runtime guard, and add hipMemsetAsync to zero the output buffer before atomic accumulation. Non-tune entry points pass KBatch=1 (no split-K) to preserve existing behavior. Code generation scripts (gen_instances.py, gen_instances_cktile.py) updated to include the new parameter in generated wrappers and manifests. Made-with: Cursor * Wire splitK from tuning CSV through production blockscale GEMM dispatch The tuning infrastructure already sweeps splitK and writes it to the CSV, but the production dispatch ignored it and hardcoded KBatch=1. Add splitK as a runtime parameter to the non-tune entry points so tuned split-K values are used without compiling the full _tune instance set. Made-with: Cursor * fix: ck_moe_stage1 split-K output buffer overflow from padding scatter The CK kernel scatters output via sorted_token_ids using: token_offset = (fused_token & 0xffffff) * topk + (fused_token >> 24) Padding entries use the sentinel value (topk << 24 | token_num), which decodes to scatter position (token_num * topk + topk) -- beyond the valid output range [0, token_num * topk). The original buffer (token_num, topk, w1.shape[1]) only has token_num * topk rows, so the padding scatter writes out of bounds, causing "HIP runtime error: invalid argument" during CUDA graph capture (e.g. DeepSeek-R1 decode with token_num=1, topk=8, block_m=16). Fix: allocate (token_num * topk + topk + 1) rows -- the exact minimum needed to absorb all padding scatter writes. After the kernel, slice only the valid [0, token_num * topk) rows for the activation. Related: #2508 Made-with: Cursor * Address PR review feedback: validate splitK, fix hipMemset stride issue, add correctness test Agent-Logs-Url: https://github.com/ROCm/aiter/sessions/e3b37b0f-e151-4935-ad89-fd72436d41e2 Co-authored-by: samremes <181322991+samremes@users.noreply.github.com> * black format * fix splitk test dimensions * Add gdn fusions * style: fix ruff F841 and black-format Triton PR files Remove unused variable in rmsnorm FP8 test ref. Apply Black to kernels, launchers, tests, and gated_delta_rule decode __init__. Made-with: Cursor * Update fused_rearrange_sigmoid_gdr.py * Update op_tests * Fix BLACK format problem * Fix black check failure * Update test_fused_rearrange_sigmoid_gdr.py * Allow callers to pass pre-allocated moe_buf to avoid output copy Add an optional `moe_buf` parameter through the moe_sorting and fused_moe call chain. When provided, the sorting kernel writes directly into the caller's buffer instead of allocating a new one, eliminating a redundant copy on the output path. Made-with: Cursor * Add moe_buf pass-through test to existing test_moe_sorting Made-with: Cursor * Replace _fast with _single_token for causal conv1d update kernels for single token decoding * Fix blck format error * Add tuned a8w8 blockscale GEMM config for Qwen3-Next-80B-A3B on MI355X Tuned 1482 shapes (TP1/TP2/TP4) for Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 on MI355X using CK + CK-TILE backends with splitK support. Depends on: - PR #2862 (CK bump for stride fix in CK-TILE blockscale) - PR #2541 (splitK support for CK/CK-TILE blockscale GEMMs) - PR #2487 (AQLayout tunable for CK-TILE blockscale 8-warp kernels) * refactor(triton): rename gated RMSNorm+FP8 op to fused_rms_gated_fp8_group_quant Colocate the gated RMSNorm + FP8 group quant path with the other fused FP8 ops. The Triton kernel is now _fused_rms_gated_fp8_group_quant_kernel in _triton_kernels/quant/fused_fp8_quant.py; the Python entry point is fused_rms_gated_fp8_group_quant in quant/fused_fp8_quant.py, with a docstring that contrasts it with fused_rms_fp8_group_quant. Remove the old rmsnorm_input_quant_fp8 module and rms_norm_input_quant_fp8 kernel file. Re-export the new symbol and helpers (get_fp8_min_max_bounds, calc_rows_per_block) from aiter.ops.triton.quant. Rename the test file to test_fused_rms_gated_fp8_group_quant.py and update test.sh. BREAKING CHANGE: rmsnorm_input_quant_fp8 is removed; use fused_rms_gated_fp8_group_quant instead. Made-with: Cursor * Retune blockscale GEMM configs to fix invalid kernelId+splitK combinations Full retune of all 1482 shapes on MI355X (gfx950, cu_num=256). Key changes: - SplitK usage dropped from 613 to 88 CK shapes (splitK > 0) - All shapes validated via --run_config (1482/1482 OK) - E2e perf: 2-8% output throughput improvement vs untuned heuristic * [Bug] pa_mqa_logits: mask OOB stores on OutLogits_buffer The gluon `_gluon_deepgemm_fp8_paged_mqa_logits_preshuffle` and `_gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx` kernels have 10 `buffer_store(ptr=OutLogits_buffer, ...)` call sites that are missing the upper-bound mask present on their sibling stores. When `context_length == max_model_len` (the last-token position in a long- context decode step), `split_context_length` is rounded UP to a `KVBlockSize` multiple at line 427 and the final prefix/suffix store then writes up to `ChunkKPerStage` float32 elements past the logical row end. With `stride_out_batch == max_model_len`, those writes cross into the next row / the next allocation, causing intermittent HIP memory-access faults on gfx950 during DeepSeek V3.2 MTP decoding. This change adds `mask=<offset> < max_model_len` to every unmasked `buffer_store` on `OutLogits_buffer` in both preshuffle kernels, matching the pattern of their already-masked neighbours. The existing `tl.where(..., -inf)` masking of the *values* is preserved; the only behavioural change is that out-of-row lanes no longer emit buffer stores. Hardware overhead is negligible: `buffer_store` with a predicate is the same SMEM descriptor path as the unmasked variant, just with a VCC mask setup. Repro + end-to-end fix evidence: see PR description. Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com> * style: fix Black formatting * style: fix Black formatting (Python 3.12 compatible) * ci: replace deprecated zmq package with pyzmq The `zmq` meta-package fails to install on some CI runners because it cannot resolve the `pyzmq` dependency. Use `pyzmq` directly, which is the actual package providing ZeroMQ bindings for Python. Fixes Triton Test Shard 7 setup failures. * ci: increase pip retries and timeout for CI reliability Set pip global retries=15 and timeout=120s in build_aiter_triton.sh to handle transient PyPI network failures on self-hosted runners. Shard 5/7 failures were caused by RemoteDisconnected during pip install. * ci: make pyzmq install non-blocking in triton test setup pyzmq is only used by aiter.dist.shm_broadcast, not by any triton test. When PyPI is unreachable on self-hosted runners, the pyzmq install failure should not block the entire CI shard. Split pyzmq into a separate pip install with || fallback so triton tests can proceed even when PyPI connectivity is degraded. * ci: retry pip install individually on batch failure When batch pip install fails (e.g., PyPI connectivity issues on self-hosted runners), retry each package individually. Only pyzmq is allowed to fail silently since it's only used by aiter.dist.shm_broadcast and not required by any CI test suite. Critical packages (pandas, einops, numpy) must still succeed. * [MLA] Fix nhead=32 non-persistent decode crash on gfx950 Commit c849fd5 ("Add bf16 MLA decode kernel for gqa_ratio=64, qseqlen=1 (non-persistent)") zeroed ptr_RP and out_16_nosplit for all non-persistent dispatch. The legacy QH16 ASM kernel used for nhead=32 (MLA_A16W16_1TG_4W_32mx1_16nx1_Coex0_Msk1_QH16.co) still writes directly to the output buffer via ptr_RP when kv_split==1. Dereferencing nullptr causes a GPU memory access fault during CUDA graph capture on MI355X (gfx950) with DeepSeek-V3.2 at TP4. Fix: - Conditionally restore ptr_RP and out_16_nosplit in the non-persistent path for legacy kernels (gqa_ratio * max_seqlen_q <= 64) while keeping nullptr for newer kernels (e.g. gqa_ratio=64). - Restore the bf16 nhead in [32,64] early-return after stage1 when num_kv_splits==1 to prevent stage2 from overwriting the kernel's direct output. Tested on MI355X TP4 with deepseek-ai/DeepSeek-V3.2 (nhead=32): - No crash during CUDA graph capture - Correct GSM8K accuracy Made-with: Cursor * revert: remove #2983 (MLA nhead=32 fix) — causes test_mla CI failures Reverting cherry-pick of #2983 from this bulk merge. The MLA nhead=32 non-persistent decode fix causes deterministic test_mla k_cache and mla_decode-absorb precision failures on CI MI35X runners (Shard 1 & 2). #2983 should go through its own PR with proper CI validation by the original author (frida-andersson). * fix: restore tuple unpack for FlyDSL fused-quant stage1 return flydsl_moe_stage1 returns (out, out_scale_sorted) when the kernel uses fused fp4/fp8 quantization. The tuple unpack logic was removed during earlier refactoring but the kernel behavior was not changed, causing fused_moe_2stages to crash with: AttributeError: 'tuple' object has no attribute 'view' Restore the unpack: detect tuple return, extract tensor and scale, handle fp4 byte-packing trim, and skip redundant Python-side requant when the kernel already produced sorted scales. * Revert leaked changes from excluded PRs #2457/#2547/#2687 in fused_moe.py - Restore import to match main: use `from aiter import fused_dynamic_mxfp4_quant_moe_sort, mxfp4_moe_sort_fwd` instead of importing from internal triton path and fp4_utils - Replace all fp4_utils.moe_mxfp4_sort() calls with mxfp4_moe_sort_fwd() using correct parameter names (cols= instead of block_size=) - Remove all moe_buf preallocated buffer additions (PR #2687 rejected): parameter defaults, if-guards, and pass-throughs in _moe_sorting_impl, moe_sorting, fused_moe, fused_moe_fake, and fused_moe_ - Fix moe_sorting_dispatch_policy type annotation: bool -> int in fused_moe_fake and fused_moe_ - Remove moe_buf pass-through test from test_moe_sorting.py - Preserve legitimate fp4_utils usage (mxfp4_to_f32, e8m0_to_f32) with local imports in stage1/stage2 fallback functions * fix: restore fp4_utils.moe_mxfp4_sort for new code paths (different output layout than mxfp4_moe_sort_fwd) * style: fix Black formatting for local imports * fix: remove rejected W4A6 QuantType remap from fused_moe_dp_shared_expert Lingpeng explicitly rejected this change (from excluded PR #2457). Reverts the QuantType.No -> per_1x32 remap for fp4x2 weights. * fix: restore silently-reverted main features from bad merge resolution aiter/fused_moe.py: - Restore to origin/main. Per sunway513's own comment, #2457 and #2547 were excluded from this bulk merge; per valarLip, #2687 was rejected. No source PR should land changes in this file. The previous state (+110/-119 vs main) was collateral damage from auto-resolved conflicts taking older sides, which silently reverted #2262 (xbf16 asm fmoe path), #2726 (FlyDSL a8w4 MoE wrapper params + fuse_quant), #2658 (CK fp8 blockscale splitk tuner support), and #2620 (mxfp4_moe_sort_hip, flagged by valarLip). op_tests/test_gemm_a8w8_blockscale.py: - Replace with a clean 3-way merge of origin/main + #2541. Now +55/-0 vs main, matching #2541's actual contribution exactly. The previous state was silently reverting #2645 (CK GEMM multi-arch + test infra: TEST_NUM_ITERS, --csv/--output args, kernel_name= param). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: remove #2464 from bulk merge per author request @xaguilar-amd asked to drop #2464 (CK MoE tuner bug fixes) from this bulk merge — they don't need it for the uplift. Verified that #2464 is the only PR in this bulk merge touching aiter/jit/core.py and aiter/utility/mp_tuner.py: the diff between the branch and origin/main on those files is exactly #2464's +9/-1 and +5/-0, with no other PR content mixed in. Restoring both files to origin/main therefore drops #2464 cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com> Co-authored-by: vecheruk-amd <vecheruk@amd.com> Co-authored-by: xaguilar-amd <xavier.aguilarfruto@amd.com> Co-authored-by: Sami Remes <samremes@amd.com> Co-authored-by: Li <chuali@amd.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: samremes <181322991+samremes@users.noreply.github.com> Co-authored-by: hellozhuo <zhuo.su@amd.com> Co-authored-by: Tres Popp <tres.popp@amd.com> Co-authored-by: Juuso Korhonen <40278371+juuso-oskari@users.noreply.github.com> Co-authored-by: Niklas Holmberg <nholmber@users.noreply.github.com> Co-authored-by: Markus Hartikainen <markus.hartikainen@amd.com> Co-authored-by: frida-andersson <fanderss@amd.com> Co-authored-by: Aliasger Zaidy <aliasger.zaidy@amd.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

sunway513 · 2026-05-05T03:36:59Z

This PR's content was bulk-merged via #3005 ([Silo] Bulk merge: kernel fixes and features, merged 2026-05-05 03:34 UTC). Please close this PR as superseded.

Tracking issue: ROCm/AI-Frameworks-Dashboard#141

…st1 (#3005) Squash-merged from main commit 2c855fb. Includes 8 atomic Silo PRs (4 bug fixes + 3 features + 1): Bug fixes: - #2457 MoE dispatch fix for Quark W4A6 (MXFP4 weights with QuantType.No) - #2464 CK MoE tuner cascading bugs - #2547 ck_moe_stage1 split-K buffer overflow (memory safety) - #2866 pa_mqa_logits OOB stores fix (memory safety) Features: - #2423 Triton optimized decode for Qwen3-Next (GDN, conv1d, fused FP8 quant) - #2541 SplitK support for CK/CKTile Block-Scale GEMMs - #2687 Allow preallocated MoE sorting buffer Conflict resolutions (3 files): - .github/workflows/aiter-test.yaml (3 blocks): took HEAD to preserve release/v0.1.13's prebuilt-image-extract CI flow. Theirs would replace it with main's inline flow which has not been validated for this release branch. - .github/workflows/vllm_benchmark.yaml (1 block): took HEAD for the same CI architecture preservation reason. - aiter/ops/triton/gated_delta_net/__init__.py (1 block): took HEAD. Theirs would expose and , but those functions exist only on main (added in a separate commit not part of this PR), so taking theirs would break import on this release branch with NameError. The new fused_rearrange_sigmoid_gdr.py wrapper that #3005 introduces is importable directly from its module path; not exporting it via __init__.py simply means library consumers must use the longer import path. Acceptable trade-off vs broken imports. 28 files changed, +5433/-47 (3 of original 31 dropped to HEAD per above). Driver: vLLM 0.21 freeze 2026-05-08 — Silo customers need these kernel fixes (especially #2547 / #2866 memory safety) on the AITER release wheel, not nightly. Verification gates added before tag: - ATOM 5-model accuracy unchanged within +/- 0.005 vs v0.1.13-rc1 - New Qwen3-Next decode codepath smoke (GDN + causal_conv1d_single_token + fused_fp8_quant must JIT-compile and produce coherent output) - Memory safety regression check on Kimi-K2.5-MXFP4 (exercises ck_moe stage1) and DeepSeek-V3.2 (exercises pa_mqa_logits) - Perf delta sample on Kimi/MiniMax/DSv3.2 c=1 + c=64 vs rc1 baseline (cherry picked from commit 2c855fb)

maeehart · 2026-05-05T17:07:09Z

This can be discarded at this moment.

@xaguilar-amd

…3-Next, pa_mqa OOB) (#3005) * fix: remap QuantType.No to per_1x32 for fp4x2 MoE weights (W4A6 support) * Fixing two cascading bugs when running the MoE tuner * Enable split-K for block-scale A8W8 CK and CKTile GEMMs Propagate the splitK parameter (as KBatch = 2^splitK) through the block-scale GEMM kernel infrastructure so that the tuning scripts can sweep split-K values to improve occupancy on small-M shapes. CK path: add KBatch parameter to gemm_a8w8_blockscale_impl and call SetKBatch on the device argument. The CK invoker handles output zeroing and atomic accumulation internally. CKTile path: add k_batch parameter to gemm_a8w8_blockscale_cktile_impl, remove the "split-k is not supported yet" runtime guard, and add hipMemsetAsync to zero the output buffer before atomic accumulation. Non-tune entry points pass KBatch=1 (no split-K) to preserve existing behavior. Code generation scripts (gen_instances.py, gen_instances_cktile.py) updated to include the new parameter in generated wrappers and manifests. Made-with: Cursor * Wire splitK from tuning CSV through production blockscale GEMM dispatch The tuning infrastructure already sweeps splitK and writes it to the CSV, but the production dispatch ignored it and hardcoded KBatch=1. Add splitK as a runtime parameter to the non-tune entry points so tuned split-K values are used without compiling the full _tune instance set. Made-with: Cursor * fix: ck_moe_stage1 split-K output buffer overflow from padding scatter The CK kernel scatters output via sorted_token_ids using: token_offset = (fused_token & 0xffffff) * topk + (fused_token >> 24) Padding entries use the sentinel value (topk << 24 | token_num), which decodes to scatter position (token_num * topk + topk) -- beyond the valid output range [0, token_num * topk). The original buffer (token_num, topk, w1.shape[1]) only has token_num * topk rows, so the padding scatter writes out of bounds, causing "HIP runtime error: invalid argument" during CUDA graph capture (e.g. DeepSeek-R1 decode with token_num=1, topk=8, block_m=16). Fix: allocate (token_num * topk + topk + 1) rows -- the exact minimum needed to absorb all padding scatter writes. After the kernel, slice only the valid [0, token_num * topk) rows for the activation. Related: #2508 Made-with: Cursor * Address PR review feedback: validate splitK, fix hipMemset stride issue, add correctness test Agent-Logs-Url: https://github.com/ROCm/aiter/sessions/e3b37b0f-e151-4935-ad89-fd72436d41e2 Co-authored-by: samremes <181322991+samremes@users.noreply.github.com> * black format * fix splitk test dimensions * Add gdn fusions * style: fix ruff F841 and black-format Triton PR files Remove unused variable in rmsnorm FP8 test ref. Apply Black to kernels, launchers, tests, and gated_delta_rule decode __init__. Made-with: Cursor * Update fused_rearrange_sigmoid_gdr.py * Update op_tests * Fix BLACK format problem * Fix black check failure * Update test_fused_rearrange_sigmoid_gdr.py * Allow callers to pass pre-allocated moe_buf to avoid output copy Add an optional `moe_buf` parameter through the moe_sorting and fused_moe call chain. When provided, the sorting kernel writes directly into the caller's buffer instead of allocating a new one, eliminating a redundant copy on the output path. Made-with: Cursor * Add moe_buf pass-through test to existing test_moe_sorting Made-with: Cursor * Replace _fast with _single_token for causal conv1d update kernels for single token decoding * Fix blck format error * Add tuned a8w8 blockscale GEMM config for Qwen3-Next-80B-A3B on MI355X Tuned 1482 shapes (TP1/TP2/TP4) for Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 on MI355X using CK + CK-TILE backends with splitK support. Depends on: - PR #2862 (CK bump for stride fix in CK-TILE blockscale) - PR #2541 (splitK support for CK/CK-TILE blockscale GEMMs) - PR #2487 (AQLayout tunable for CK-TILE blockscale 8-warp kernels) * refactor(triton): rename gated RMSNorm+FP8 op to fused_rms_gated_fp8_group_quant Colocate the gated RMSNorm + FP8 group quant path with the other fused FP8 ops. The Triton kernel is now _fused_rms_gated_fp8_group_quant_kernel in _triton_kernels/quant/fused_fp8_quant.py; the Python entry point is fused_rms_gated_fp8_group_quant in quant/fused_fp8_quant.py, with a docstring that contrasts it with fused_rms_fp8_group_quant. Remove the old rmsnorm_input_quant_fp8 module and rms_norm_input_quant_fp8 kernel file. Re-export the new symbol and helpers (get_fp8_min_max_bounds, calc_rows_per_block) from aiter.ops.triton.quant. Rename the test file to test_fused_rms_gated_fp8_group_quant.py and update test.sh. BREAKING CHANGE: rmsnorm_input_quant_fp8 is removed; use fused_rms_gated_fp8_group_quant instead. Made-with: Cursor * Retune blockscale GEMM configs to fix invalid kernelId+splitK combinations Full retune of all 1482 shapes on MI355X (gfx950, cu_num=256). Key changes: - SplitK usage dropped from 613 to 88 CK shapes (splitK > 0) - All shapes validated via --run_config (1482/1482 OK) - E2e perf: 2-8% output throughput improvement vs untuned heuristic * [Bug] pa_mqa_logits: mask OOB stores on OutLogits_buffer The gluon `_gluon_deepgemm_fp8_paged_mqa_logits_preshuffle` and `_gluon_deepgemm_fp8_paged_mqa_logits_preshuffle_varctx` kernels have 10 `buffer_store(ptr=OutLogits_buffer, ...)` call sites that are missing the upper-bound mask present on their sibling stores. When `context_length == max_model_len` (the last-token position in a long- context decode step), `split_context_length` is rounded UP to a `KVBlockSize` multiple at line 427 and the final prefix/suffix store then writes up to `ChunkKPerStage` float32 elements past the logical row end. With `stride_out_batch == max_model_len`, those writes cross into the next row / the next allocation, causing intermittent HIP memory-access faults on gfx950 during DeepSeek V3.2 MTP decoding. This change adds `mask=<offset> < max_model_len` to every unmasked `buffer_store` on `OutLogits_buffer` in both preshuffle kernels, matching the pattern of their already-masked neighbours. The existing `tl.where(..., -inf)` masking of the *values* is preserved; the only behavioural change is that out-of-row lanes no longer emit buffer stores. Hardware overhead is negligible: `buffer_store` with a predicate is the same SMEM descriptor path as the unmasked variant, just with a VCC mask setup. Repro + end-to-end fix evidence: see PR description. Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com> * style: fix Black formatting * style: fix Black formatting (Python 3.12 compatible) * ci: replace deprecated zmq package with pyzmq The `zmq` meta-package fails to install on some CI runners because it cannot resolve the `pyzmq` dependency. Use `pyzmq` directly, which is the actual package providing ZeroMQ bindings for Python. Fixes Triton Test Shard 7 setup failures. * ci: increase pip retries and timeout for CI reliability Set pip global retries=15 and timeout=120s in build_aiter_triton.sh to handle transient PyPI network failures on self-hosted runners. Shard 5/7 failures were caused by RemoteDisconnected during pip install. * ci: make pyzmq install non-blocking in triton test setup pyzmq is only used by aiter.dist.shm_broadcast, not by any triton test. When PyPI is unreachable on self-hosted runners, the pyzmq install failure should not block the entire CI shard. Split pyzmq into a separate pip install with || fallback so triton tests can proceed even when PyPI connectivity is degraded. * ci: retry pip install individually on batch failure When batch pip install fails (e.g., PyPI connectivity issues on self-hosted runners), retry each package individually. Only pyzmq is allowed to fail silently since it's only used by aiter.dist.shm_broadcast and not required by any CI test suite. Critical packages (pandas, einops, numpy) must still succeed. * [MLA] Fix nhead=32 non-persistent decode crash on gfx950 Commit c849fd5 ("Add bf16 MLA decode kernel for gqa_ratio=64, qseqlen=1 (non-persistent)") zeroed ptr_RP and out_16_nosplit for all non-persistent dispatch. The legacy QH16 ASM kernel used for nhead=32 (MLA_A16W16_1TG_4W_32mx1_16nx1_Coex0_Msk1_QH16.co) still writes directly to the output buffer via ptr_RP when kv_split==1. Dereferencing nullptr causes a GPU memory access fault during CUDA graph capture on MI355X (gfx950) with DeepSeek-V3.2 at TP4. Fix: - Conditionally restore ptr_RP and out_16_nosplit in the non-persistent path for legacy kernels (gqa_ratio * max_seqlen_q <= 64) while keeping nullptr for newer kernels (e.g. gqa_ratio=64). - Restore the bf16 nhead in [32,64] early-return after stage1 when num_kv_splits==1 to prevent stage2 from overwriting the kernel's direct output. Tested on MI355X TP4 with deepseek-ai/DeepSeek-V3.2 (nhead=32): - No crash during CUDA graph capture - Correct GSM8K accuracy Made-with: Cursor * revert: remove #2983 (MLA nhead=32 fix) — causes test_mla CI failures Reverting cherry-pick of #2983 from this bulk merge. The MLA nhead=32 non-persistent decode fix causes deterministic test_mla k_cache and mla_decode-absorb precision failures on CI MI35X runners (Shard 1 & 2). #2983 should go through its own PR with proper CI validation by the original author (frida-andersson). * fix: restore tuple unpack for FlyDSL fused-quant stage1 return flydsl_moe_stage1 returns (out, out_scale_sorted) when the kernel uses fused fp4/fp8 quantization. The tuple unpack logic was removed during earlier refactoring but the kernel behavior was not changed, causing fused_moe_2stages to crash with: AttributeError: 'tuple' object has no attribute 'view' Restore the unpack: detect tuple return, extract tensor and scale, handle fp4 byte-packing trim, and skip redundant Python-side requant when the kernel already produced sorted scales. * Revert leaked changes from excluded PRs #2457/#2547/#2687 in fused_moe.py - Restore import to match main: use `from aiter import fused_dynamic_mxfp4_quant_moe_sort, mxfp4_moe_sort_fwd` instead of importing from internal triton path and fp4_utils - Replace all fp4_utils.moe_mxfp4_sort() calls with mxfp4_moe_sort_fwd() using correct parameter names (cols= instead of block_size=) - Remove all moe_buf preallocated buffer additions (PR #2687 rejected): parameter defaults, if-guards, and pass-throughs in _moe_sorting_impl, moe_sorting, fused_moe, fused_moe_fake, and fused_moe_ - Fix moe_sorting_dispatch_policy type annotation: bool -> int in fused_moe_fake and fused_moe_ - Remove moe_buf pass-through test from test_moe_sorting.py - Preserve legitimate fp4_utils usage (mxfp4_to_f32, e8m0_to_f32) with local imports in stage1/stage2 fallback functions * fix: restore fp4_utils.moe_mxfp4_sort for new code paths (different output layout than mxfp4_moe_sort_fwd) * style: fix Black formatting for local imports * fix: remove rejected W4A6 QuantType remap from fused_moe_dp_shared_expert Lingpeng explicitly rejected this change (from excluded PR #2457). Reverts the QuantType.No -> per_1x32 remap for fp4x2 weights. * fix: restore silently-reverted main features from bad merge resolution aiter/fused_moe.py: - Restore to origin/main. Per sunway513's own comment, #2457 and #2547 were excluded from this bulk merge; per valarLip, #2687 was rejected. No source PR should land changes in this file. The previous state (+110/-119 vs main) was collateral damage from auto-resolved conflicts taking older sides, which silently reverted #2262 (xbf16 asm fmoe path), #2726 (FlyDSL a8w4 MoE wrapper params + fuse_quant), #2658 (CK fp8 blockscale splitk tuner support), and #2620 (mxfp4_moe_sort_hip, flagged by valarLip). op_tests/test_gemm_a8w8_blockscale.py: - Replace with a clean 3-way merge of origin/main + #2541. Now +55/-0 vs main, matching #2541's actual contribution exactly. The previous state was silently reverting #2645 (CK GEMM multi-arch + test infra: TEST_NUM_ITERS, --csv/--output args, kernel_name= param). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: remove #2464 from bulk merge per author request @xaguilar-amd asked to drop #2464 (CK MoE tuner bug fixes) from this bulk merge — they don't need it for the uplift. Verified that #2464 is the only PR in this bulk merge touching aiter/jit/core.py and aiter/utility/mp_tuner.py: the diff between the branch and origin/main on those files is exactly #2464's +9/-1 and +5/-0, with no other PR content mixed in. Restoring both files to origin/main therefore drops #2464 cleanly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Signed-off-by: Markus Hartikainen <markus.hartikainen@amd.com> Co-authored-by: vecheruk-amd <vecheruk@amd.com> Co-authored-by: xaguilar-amd <xavier.aguilarfruto@amd.com> Co-authored-by: Sami Remes <samremes@amd.com> Co-authored-by: Li <chuali@amd.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: samremes <181322991+samremes@users.noreply.github.com> Co-authored-by: hellozhuo <zhuo.su@amd.com> Co-authored-by: Tres Popp <tres.popp@amd.com> Co-authored-by: Juuso Korhonen <40278371+juuso-oskari@users.noreply.github.com> Co-authored-by: Niklas Holmberg <nholmber@users.noreply.github.com> Co-authored-by: Markus Hartikainen <markus.hartikainen@amd.com> Co-authored-by: frida-andersson <fanderss@amd.com> Co-authored-by: Aliasger Zaidy <aliasger.zaidy@amd.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Fixing two cascading bugs when running the MoE tuner

feb21bc

xaguilar-amd marked this pull request as ready for review March 25, 2026 13:43

xaguilar-amd requested a review from a team March 25, 2026 13:43

valarLip requested a review from amd-ruitang3 March 25, 2026 14:06

sunway513 mentioned this pull request May 1, 2026

[Silo] Bulk merge: kernel fixes and features (SplitK, MoE fixes, Qwen3-Next, pa_mqa OOB) #3005

Merged

3 tasks

maeehart closed this May 5, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fixing two cascading bugs when running the CK MoE tuner#2464

Fixing two cascading bugs when running the CK MoE tuner#2464
xaguilar-amd wants to merge 1 commit into
ROCm:mainfrom
xaguilar-amd:xaguilar-amd/fix-ck-moe_tuning

xaguilar-amd commented Mar 25, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Mar 25, 2026

Uh oh!

sunway513 commented May 5, 2026

Uh oh!

maeehart commented May 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

xaguilar-amd commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Root causes

1 — module_moe_asm.so was compiled without torch linkage (aiter/jit/core.py)

2 — Failed tasks left result_dict empty, causing IndexError (aiter/utility/mp_tuner.py)

How the two bugs interact

Testing

Notes for reviewers

Uh oh!

github-actions Bot commented Mar 25, 2026

🏷️ CI Guide

Uh oh!

sunway513 commented May 5, 2026

Uh oh!

maeehart commented May 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

xaguilar-amd commented Mar 25, 2026 •

edited

Loading

1 — `module_moe_asm.so` was compiled without torch linkage (`aiter/jit/core.py`)

2 — Failed tasks left `result_dict` empty, causing `IndexError` (`aiter/utility/mp_tuner.py`)