Skip to content

feat(EPv2): hybrid backend — HIP dispatch + FlyDSL combine - #563

Open
kawhil-amd wants to merge 3 commits into
ROCm:mainfrom
kawhil-amd:feat/epv2-hybrid-backend
Open

feat(EPv2): hybrid backend — HIP dispatch + FlyDSL combine#563
kawhil-amd wants to merge 3 commits into
ROCm:mainfrom
kawhil-amd:feat/epv2-hybrid-backend

Conversation

@kawhil-amd

Copy link
Copy Markdown
Contributor

No description provided.

kawhil-amd and others added 3 commits August 14, 2026 08:22
Split KernelSet's stages_in_kernel / self_resets_counters into per-leg
flags (dispatch_resets_counters, combine_resets_counters,
combine_stages_in_kernel) so the base class can drive each leg with the
correct reset and staging semantics independently.

Add hybrid_backend.py: EpDispatchCombineOpHybrid uses HIP/JIT
EpDispatchPlan for the dispatch leg and FlyDSL make_combine for the
gather-combine leg, sharing one SymmArena.  Selectable via
MORI_V2_KERNEL_BACKEND=hybrid or cfg.kernel_backend="hybrid".

Tuning cross-picks dispatch geometry from hip_tuning_configs and combine
geometry from tuning_configs, merged via merge_buckets (renamed from
_merge for importability).

Backend parity test now covers all three backends (flydsl, hip, hybrid).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The hybrid backend accepted fp4/fp8 combine configs silently (no gate in
_unsupported). Add an explicit combine_dtype check matching the bf16/fp32
constraint. Also update the test comment from "2 backends" to "3 arenas".

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
_arch/_is1250 were computed but never used: the HIP dispatch plan picks
its portable-vs-TDM body at render time, and the gfx1250 combine fan
buffer is irrelevant because hybrid combine is FlyDSL, not the TDM kernel.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant