Skip to content

Annier.rocmfp4 support - #86

Merged
jimw567 merged 3 commits into
gfx11from
annier.rocmfp4-support
Aug 16, 2026
Merged

jimw567 merged 3 commits into
gfx11from
annier.rocmfp4-support

Conversation

@Annieren

Copy link
Copy Markdown

Overview

Integrated RCOmfp4 support from https://github.com/charlie12345/rocmfp4-llama

Additional information

Requirements

  • I have read and agree with the contributing guidelines
  • AI usage disclosure: Yes, it helps understanding and testing.

Annie Ren and others added 3 commits July 31, 2026 12:10
Port the self-contained ROCmFP4 Codebook10 dequant module into ggml/rocmfp4/
and wire the new GGML_TYPE_Q4_0_ROCMFP4(=100)/_FAST(=101) types through ggml
core, the CPU backend, and the HIP dequant path so the Strix qwen35moe GGUF
loads and runs correctly on gfx1151 (no native FP4). Matmul falls back to the
cuBLAS/hipBLAS dequant path (correct but slow); MMVQ/MMQ deliberately left off.

Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Add vec_dot_rocmfp4[_fast]_q8_1 device fns (vecdotq.cuh) and wire rocmfp4 into
the mmvq type switch (get_vec_dot_q_cuda / get_vdr_mmvq / mul_mat_vec_q_switch_
type), and enable the MMVQ path (incl. MoE mul_mat_id) for rocmfp4. Decode goes
from ~11 t/s (cuBLAS-dequant) to ~75 t/s on gfx1151; output is bit-exact vs M1.

Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
@Annieren
Annieren requested a review from jimw567 August 16, 2026 21:35

@jimw567 jimw567 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

looks good

@jimw567
jimw567 merged commit d00944d into gfx11 Aug 16, 2026
9 checks passed
@jimw567
jimw567 deleted the annier.rocmfp4-support branch August 16, 2026 23:34
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.

2 participants