Repository navigation
cleanup(rocm): trim dead W4A8 explore sweep from production copy - #41
Closed
BlivionIaG wants to merge 1 commit into
Closed
BlivionIaG wants to merge 1 commit into
BlivionIaG wants to merge 1 commit into
Conversation
The gfx1030 production GEMM instantiated nine explore configs times three group sizes, but w4a8_gemm_rdna2 only ever called find_config(8). Keep that config (groups 32/64/128) and drop the unused sweep, probe kernel, and shape loggers. Co-authored-by: Grok 4.7 <noreply@x.ai> Signed-off-by: Cursor Agent <cursoragent@cursor.com>
This was referenced Oct 7, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Trim the production gfx1030 W4A8 path down to the one config that actually launches.
w4a8_gemm_rdna2calledfind_config(8), which isa8_lds_k32_ag(256 threads, NPT 4, K_STEP 32, M_TILE 8, LDS A, per-group activation scales). The production TU still instantiated the other eight explore configs at group sizes 32/64/128, plus an unlaunched probe kernel and two shape loggers. This keepsa8_lds_k32_agand those three group sizes, and deletes the rest.Not a duplicate: no open PR on
opengfx1030/vllm-rdnatrims this sweep. Left untouched:docs/explore/,csrc/rocm/explore/(not present in this tree), and the open territories in #25, #2, #33, #37, and #40.AI assistance was used (Cursor Grok 4.7). A human should review every changed line before this leaves draft.
What stayed
a8_lds_k32_agfor group sizes 32, 64, and 128, viakLaunchA8LdsK32Ag.TORCH_WARN_ONCEon the dense path (config a8_lds_k32_ag) and the MoE path (config moe_a8_k32_ag).w4a8_sdot4_rdna2.cuh.VLLM_ROCM_EXT_SRC(CMakeLists.txt:csrc/rocm/w4a8_sdot4_rdna2.cu,csrc/rocm/moe_w4a8_rdna2.cu).What went
find_configtable. Only the liveCfg<256, 4, 32, 8, {32,64,128}, kLds, true>launches remain.w4a8_probe_kernelandW4A8_EXPLORE_INSTANTIATE_GEMM./tmp/pr9_branchchangelog in the.cu.w4a8_log_shape/VLLM_RDNA2_W4A8_DEBUG,w4a8_moe_log_shape/VLLM_RDNA2_W4A8_MOE_DEBUG). The MoE workspace cache is unchanged.Diff: 5 files, +12 / −223.
Test Plan
rg -n 'W4A8_EXPLORE_CONFIGS|W4A8_EXPLORE_INSTANTIATE_GEMM|w4a8_probe_kernel|w4a8_log_shape|w4a8_debug_enabled|w4a8_moe_log_shape|w4a8_moe_debug_enabled|\bfind_config\b' .venv/bin/python -m pytest tests/kernels/quantization/test_rdna2_w4a8.py tests/kernels/quantization/test_rdna2_moe_w4a8.py -vTest Result
Symbol sweep is empty (no remaining references).
Kernel tests were not run. This environment has no ROCm stack (
hipcc//opt/rocmabsent) and no GPU, so the gfx1030-gated tests cannot collect against a built_rocm_C. The usual skip isrequires gfx1030 with w4a8_act_quant_rdna2 / w4a8_gemm_rdna2 built in.No model eval: launch parameters of the live config are unchanged, and the deleted configs were never dispatched.