Repository navigation
feat(hip/gfx906): DS4 SDOT4 index score, FP16 DSpark FC16, row-local GLM prefill, and Q8_0 MM3 GGUF parity - #889
Draft
nicholas631 wants to merge 4 commits into
Conversation
nicholas631
force-pushed
the
feat/hip-gfx906-minimax-ep4-ds4-dspark8
branch
from
November 14, 2026 04:29
3606727 to
5be320c
Compare
nicholas631
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Rebased cleanly onto Project Maya
v1.1.63(6651b5b) with zero host-specific recipes or benchmark artifacts.Modified Files
CMakeLists.txt: Registerglm_prefill_row_local_test.docs/SETTINGS.md: DocumentMAYA_GLM_PREFILL_ROW_LOCAL.src/core/glm_fast_path.cu,src/core/glm_fast_state.hpp,src/core/glm_prefill.cu,src/core/glm_prefill_row_local.hpp: Row-local GLM prefill path (MAYA_GLM_PREFILL_ROW_LOCAL) avoiding large shared-memory staging on wave64 gfx906.src/ds4/ds4_kernels.cu: AddDS4_INDEX_SDOT4=1index_score_sdot4_kernelfor__gfx906__guarded tohd == 128.src/kernels/cuda/iq_kernels.cu: AllowQ8_0down-projection innative_expert_grouped.src/mm3/mm3_dspark.cu: Add gfx906 FP16 DSpark FC16 (MAYA_MM3_DSPARK_FC16) path.src/mm3/mm3_gguf.cu,tests/mm3_gguf_test.cu: AddQ8_0MM3 GGUF parity support.tests/glm_prefill_row_local_test.cpp: Unit + numerical parity test for row-local IQ3_S gate/up and IQ4_XS down prefill.Validation & Numerical Parity
./build-gfx906/glm_prefill_row_local_test,./build-gfx906/ds4_gguf_test,./build-gfx906/iq_mmvq_parity_testhip_intrinsics,ds4_gguf_test,iq_mmvq_parity_test, andglm_prefill_row_local_test(PASS).sm_86,gfx1201) unaffected behind__gfx906__/ opt-in runtime guards.