Repository navigation

All pull requests

Labels (10)Milestones (0)New pull request

Open

(21)

Closed

(129)

hrx: publish decode attention output before Q8 packing
#154AaronStGeorge opened Oct 21, 2026Collaborator
Add Q4_0 direct fp32 activation kernel support
#152Annieren opened Oct 20, 2026
rdna35 prefill optimizations other than mmq
#129liangliangchang opened Oct 1, 2026
ggml-hrx: fix empty views and improve long-context decode
#114jammm opened Sep 23, 2026
llama-server : make chat logprobs response OpenAI-compatible
#113Json288 opened Sep 22, 2026Collaborator
7
Fix hang on exit and build issues on Windows
#105pschlan-amd opened Sep 18, 2026
Fix: Return 'invalid_prompt' error type for invalid token IDs
#100amd-callumm opened Sep 8, 2026
11
cuda/hip: optional wave64 for the quantized mat-vec kernels on RDNA
#89mgehre-amd opened Aug 17, 2026Collaborator
q4_K mat-vec tuning for gfx1151
#84mgehre-amd opened Aug 13, 2026Collaborator
CUDA: dispatch the chunked gated_delta_net in CU mode when it fills the device
#79roberteg16 opened Jul 30, 2026
4
llama: let the backend pick n_ubatch, and default to 2048 on RDNA3.5
#77roberteg16 opened Jul 29, 20263 tasks done
tools: add mmq-tune, an MMQ tile-width autotuning harness
#76roberteg16 opened Jul 29, 20263 tasks done
ggml-cuda: f32 tall-skinny GEMM kernel for gfx1151 (RDNA3.5)
#74roberteg16 opened Jul 28, 2026
HRX API updates
#73AaronStGeorge opened Jul 27, 2026Collaborator
feat(cuda): fuse activations and residual add into mmv f/q epilogues
#67roberteg16 opened Jul 24, 2026
tests: add MoE MMQ benchmark with routing-distribution generator
#62roberteg16 opened Jul 20, 20263 tasks
gfx1151 (Strix Halo): fuse attn_k+v into single MMVQ dispatch
#59jeffli-xilinx opened Jul 17, 20267 tasks done
16
ggml-cuda: GEMM weight row padding + one-time K-padded f16 dequant for prefill
#57roberteg16 opened Jul 17, 20264 of 5 tasks
ggml-cuda: use mma-f16 flash-attn for D=256 on RDNA3.5
#40roberteg16 opened Jul 8, 2026
1
Basic CI for hrx-v2
#27AaronStGeorge opened Jun 18, 2026Collaborator
Updates for new hrx C and CMake APIs
#26AaronStGeorge opened Jun 18, 2026Collaborator

Sunbelt Computer Software

PL/B Language Development and Support

Navigation Menu

Repository navigation

All pull requests

hrx: publish decode attention output before Q8 packing

Add Q4_0 direct fp32 activation kernel support

rdna35 prefill optimizations other than mmq

ggml-hrx: fix empty views and improve long-context decode

llama-server : make chat logprobs response OpenAI-compatible

Fix hang on exit and build issues on Windows

Fix: Return 'invalid_prompt' error type for invalid token IDs

cuda/hip: optional wave64 for the quantized mat-vec kernels on RDNA

q4_K mat-vec tuning for gfx1151

CUDA: dispatch the chunked gated_delta_net in CU mode when it fills the device

llama: let the backend pick n_ubatch, and default to 2048 on RDNA3.5

tools: add mmq-tune, an MMQ tile-width autotuning harness

ggml-cuda: f32 tall-skinny GEMM kernel for gfx1151 (RDNA3.5)

HRX API updates

feat(cuda): fuse activations and residual add into mmv f/q epilogues

tests: add MoE MMQ benchmark with routing-distribution generator

gfx1151 (Strix Halo): fuse attn_k+v into single MMVQ dispatch

ggml-cuda: GEMM weight row padding + one-time K-padded f16 dequant for prefill

ggml-cuda: use mma-f16 flash-attn for D=256 on RDNA3.5

Basic CI for `hrx-v2`

Updates for new hrx C and CMake APIs

Sunbelt Computer Software

PL/B Language Development and Support

Pull requests