fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel by karverma-amd · Pull Request #2508 · ROCm/aiter

karverma-amd · 2026-03-27T16:58:29Z

Motivation

The splitK optimization in aiter's CK MoE kernel reduces average kernel time from 44us to 22us (2x improvement). However, enabling it in aiter commit: f80d9c0 causes:

RuntimeError: HIP runtime error: invalid argument. hip_check_error.hpp: 18
in function: hip_check_error

Technical Details

The crash occurs during CUDA graph capture in vLLM when ksplit > 1 for decode batches with small token counts (token_num=1, topk=8).

Root Cause

Buffer overflow in tmp_out allocation in ck_moe_stage1() (/app/aiter/aiter/fused_moe.py line ~1638).

The Mismatch

When splitK is active (KBatch > 1), the CK C++ kernel calls hipMemsetAsync to zero-initialize the output buffer for atomic accumulation. The size it zeroes is:

M * N * sizeof(float) * 2

where M = sorted_size = min(token_num * topk * block_m, sorted_token_ids.shape[0]).

But Python allocated tmp_out as:

torch.zeros((token_num, topk, w1.shape[1]), dtype=fp32)

For DeepSeek V3 decode with token_num=1, topk=8, block_m=16:

Python buffer: 1 * 8 * 4096 * 4 = 131,072 bytes (128 KB)
C++ hipMemsetAsync: 128 * 2048 * 4 * 2 = 2,097,152 bytes (2 MB)
Overflow: 16x beyond the allocated tensor

C++ Code Reference

The overflow originates in device_moe_gemm_blockscale.hpp (CK):

if(arg.KBatch > 1)
    hipGetErrorString(hipMemsetAsync(arg.p_c_grid, 0,
        arg.M * arg.N * sizeof(CDataType) * (IsInputGemm && IsSplitK ? 2 : 1),
        stream_config.stream_id_));

Fix

Replace the tmp_out allocation to match the kernel's expected buffer size:

# BEFORE (buggy):
tmp_out = torch.zeros(
    (token_num, topk, w1.shape[1]), dtype=dtypes.fp32, device=out.device
)

# AFTER (fixed):
sorted_size = min(token_num * topk * block_m, sorted_token_ids.shape[0])
tmp_out = torch.zeros(
    (sorted_size, w1.shape[1]), dtype=dtypes.fp32, device=out.device
)

Test Result

DeepSeek-R1-0528 with aiter commit - f80d9c00b9c8 runtime error is resolved.

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

github-actions · 2026-03-27T16:59:22Z

🏷️ 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 2508 --add-label <label>

ChuanLi1101

Looks good with minor suggestions. Pre Approved for speed up the deliveries.

ChuanLi1101 · 2026-03-31T05:31:39Z

 ):
    token_num = hidden_states.shape[0]
    is_splitk = quant_type is aiter.QuantType.per_1x128 and splitk > 1
-    tmp_out = (


cktile_moe_stage1 has the identical undersized buffer pattern for split_k > 1. It calls aiter.moe_cktile2stages_gemm1 which is a different kernel, but if that kernel has the same hipMemsetAsync behavior, it would have the same overflow. The TODO at line 1723 (# TODO: support fp32 splitk) suggests this path may not be actively used yet, but it's worth either applying the same fix or adding a guard/comment.

valarLip · 2026-03-31T05:43:36Z

+    if is_splitk:
+        # CK splitK kernel hipMemsetAsync zeros sorted_size * w1.shape[1] floats
+        sorted_size = min(token_num * topk * block_m, sorted_token_ids.shape[0])
+        tmp_out = torch.zeros(


not good to me, looks like we do zeros twice, 1 here, 1 ck, we need remove one

This ROCm/rocm-libraries#5225 already addressed the proper allocation from the C++ side. I have swap to torch.empty to avoid the double zeroing #2551 (created new PR because i cannot edit current branch)
@ChuanLi1101 i see you also created another PR but I think the double zeroing still there?

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

Base image: rocm/vllm-dev:base_custom_rocm_7.2.1_torch_triton_0330_vllm018 Patches applied: - AITER SplitK bug fix (ROCm/aiter#2508) - vLLM persistent MLA kernel (vllm-project/vllm#36574) - vLLM fused AllReduce+RMSNorm (vllm-project/vllm#37891) Made-with: Cursor

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

…ble-zeroing (#2551) * fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel * Update fused_moe.py * tmp_out to use torch.empty vs. torch.zeros to avoid double zeroing Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> * tighten valid_out slice: drop redundant .contiguous() Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> * restore .view(dtypes.fp32) on valid_out for silu_and_mul/gelu_and_mul --------- Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> Co-authored-by: Karan Verma <karan.verma@amd.com>

karverma-amd · 2026-03-31T15:55:47Z

Alternative PR merged - #2551

…ble-zeroing (#2551) * fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel * Update fused_moe.py * tmp_out to use torch.empty vs. torch.zeros to avoid double zeroing Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> * tighten valid_out slice: drop redundant .contiguous() Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> * restore .view(dtypes.fp32) on valid_out for silu_and_mul/gelu_and_mul --------- Signed-off-by: rbrugaro <rita.brugarolasbrufau@amd.com> Co-authored-by: Karan Verma <karan.verma@amd.com>

…r shared experts Switch the MXFP4 a4w4 (fp4x2 activations + fp4x2 weights) MoE path from CKTile to CK JIT kernels (ck_moe_stage1 / ck_moe_stage2_fwd) and fix several dispatch bugs that caused crashes on MiniMax-M2.1-MXFP4 at TP=8. Root cause 1 -- wrong stage1 kernel for a4w4: cktile_moe_stage1 dispatches fp4x2 activations through the fp8 pipeline (F8xMXF4FlatmmPipelineAGmemBGmemCRegV1), which misinterprets packed fp4 as fp8 and produces garbage. Switch to ck_moe_stage1 which uses the correct DeviceMoeGemmMXBPreShuffle<FP4X2, FP4X2, ...> instances from the JIT-compiled module_moe_ck2stages_fp4x2_fp4x2_preshuffle_on module. Root cause 2 -- ksplit>1 steals a4w4 cases for shared experts: MiniMax has shared expert layers with inter_dim=256 (per TP=8 rank). get_ksplit() returns 2 for these, and the "ksplit > 1 and is_shuffled" elif fired before the a4w4 elif because it only checked q_dtype_w (not q_dtype_a). This routed fp4x2 activations through cktile_moe_stage1 (bf16 output) into stage2 with the b16_fp4x2_preshuffle_on module, whose heuristic dispatch has no bf16xfp4x2 kernels -> crash. Fix: add "q_dtype_a not in [dtypes.fp4x2]" guard. Root cause 3 -- ksplit must be 0 for a4w4 in fused_moe_2stages: When metadata.ksplit > 1, fused_moe_2stages skips a1 fp4x2 quantization and leaves a2_scale=None (bf16 activations into stage2). Force ksplit=0 in the a4w4 MOEMetadata so activation quantization always runs. Additional fixes: - cktile_moe_stage1 split_k>1 buffer overflow: allocate tmp_out with sorted_size rows (= max_num_tokens_padded) instead of token_num*topk rows, mirroring the fix from ck_moe_stage1 (#2508). - cktile_moe_stage1 split_k>1 stage2: use ck_moe_stage2_fwd (JIT CK) instead of cktile_moe_stage2 for correct preshuffle dispatch. - shuffle_weight_a16w4: set is_shuffled=True on returned tensor. - Add e8m0_unshuffle utility (inverse of e8m0_shuffle). - test_moe_2stage: add a4w4 branch using shuffle_weight_a16w4(gate_up=False) matching the CK JIT kernel expected layout. Tested: MiniMax-M2.1-MXFP4 TP=8 server starts, warmup completes (all cuda_graph_capture_sizes including 48), inference produces correct output. Files changed: aiter/fused_moe.py aiter/ops/shuffle.py aiter/utility/fp4_utils.py op_tests/test_moe_2stage.py

@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>

@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>

fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel

f7ddc47

karverma-amd requested a review from a team March 27, 2026 16:58

karverma-amd added the bug Something isn't working label Mar 30, 2026

karverma-amd requested a review from valarLip March 30, 2026 16:33

karverma-amd added 3 commits March 30, 2026 11:33

Merge branch 'main' into main

3f50525

Update fused_moe.py

857c9d2

Merge branch 'main' into main

5063b5b

ChuanLi1101 self-assigned this Mar 31, 2026

ChuanLi1101 approved these changes Mar 31, 2026

View reviewed changes

valarLip reviewed Mar 31, 2026

View reviewed changes

ChuanLi1101 mentioned this pull request Mar 31, 2026

fix: ck_moe_stage1 split-K buffer overflow from padding scatter (alternative to #2508) #2547

Closed

3 tasks

karverma-amd closed this Mar 31, 2026

sogalin mentioned this pull request Apr 29, 2026

[Issue]: GPT-OSS-120b generated rubbish results with v0.1.12.post1 in SGL #2960

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel #2508

fix: align ck_moe_stage1 split-K tmp_out buffer with CK kernel #2508
karverma-amd wants to merge 4 commits into
ROCm:mainfrom
karverma-amd:main

karverma-amd commented Mar 27, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Mar 27, 2026

Uh oh!

ChuanLi1101 left a comment

Uh oh!

ChuanLi1101 Mar 31, 2026

Uh oh!

valarLip Mar 31, 2026

Uh oh!

rbrugaro-amd Mar 31, 2026 •

edited

Loading

Uh oh!

karverma-amd commented Mar 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

karverma-amd commented Mar 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Root Cause

The Mismatch

C++ Code Reference

Fix

Test Result

Submission Checklist

Uh oh!

github-actions Bot commented Mar 27, 2026

🏷️ CI Guide

Uh oh!

ChuanLi1101 left a comment

Choose a reason for hiding this comment

Uh oh!

ChuanLi1101 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

valarLip Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

rbrugaro-amd Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

karverma-amd commented Mar 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

karverma-amd commented Mar 27, 2026 •

edited

Loading

rbrugaro-amd Mar 31, 2026 •

edited

Loading