[ROCm][Compile] Fuse RMSNorm + MXFP4 quant via AITER Triton kernels (DeepSeek-R1)#44437

Draft

shantipriya-amd wants to merge 14 commits into

vllm-project:mainfrom

shantipriya-amd:feat/uplift-dsv3/pr1-register-env-vars

Commits on Jun 4, 2026

feat(rocm): register VLLM_ROCM_USE_AITER_TRITON_FUSED_RMSNORM_FP4_QUANT and FUSED_ROPE_ZEROS_KV_CACHE env vars

shantipriya-amd
and
GitHub Copilot
committed
feat(rocm): rename to FUSION_* namespace, wire _aiter_ops F2/F3, add F3 Triton dispatch in mla.py

shantipriya-amd
and
GitHub Copilot
committed
fix(rocm): correct q_out docstring shape in fused_rope_and_mla_kv_cache_write
shantipriya-amd
committed
test(rocm): TC-1.x–TC-4.x fusion flag tests for F2/F3 dispatch
shantipriya-amd
committed
test(rocm): extend TC-3.x to cover DeepSeek V2-Lite (num_q_heads=16)
shantipriya-amd
committed
feat(rocm): add MXFP4 fusion patterns + ops for RMSNorm+MXFP4-quant (F2)

shantipriya-amd
and
GitHub Copilot
committed
refactor(rocm): remove AR+MXFP4 fusion ops — defer to follow-on PR
shantipriya-amd
committed
test(rocm): remove AR+MXFP4 test stubs from test files
shantipriya-amd
committed
Merge branch 'main' into feat/uplift-dsv3/pr1-register-env-vars
shantipriya-amd
authored
fix(mxfp4): add _pattern_replacements tracking, INFO logging, fix maybe_mark_dynamic
shantipriya-amd
committed