Repository navigation

Bringing in the v3 version of the rfc - #102

Merged

rsuderman merged 50 commits into

AMD-Ecosystem:hrx-graph-develop-v2from

rsuderman:hrx-rfc-v3

Sep 16, 2026

Merged

Bringing in the v3 version of the rfc#102
rsuderman merged 50 commits into
AMD-Ecosystem:hrx-graph-develop-v2from
rsuderman:hrx-rfc-v3

Commits on Sep 3, 2026
- Generalize routed SwiGLU f16 WMMA dispatch
  rsuderman
  committed
  70c3436
- hrx: use merge mode for kernel corpus linking
  rsuderman
  committed
  f4bd4f1
- hrx: support gemma get rows vocab size
  rsuderman
  committed
  4127b66
- hrx: support bf16 get rows
  rsuderman
  committed
  c5538e7
- hrx: support wider flash attention heads
  rsuderman
  committed
  66425aa
- hrx: add Q5_K common datatype support
  rsuderman
  committed
  37830b8
- Add common IQ4_XS HRX dequant support
  rsuderman
  committed
  5118d64
- Add common HRX Q3_K dequant support
  rsuderman
  committed
  82461d3
- Add common IQ4_NL matmul dequant support
  rsuderman
  committed
  1245a51
- ggml-hrx: add common IQ3_S matmul dequant support
  rsuderman
  committed
  45d6e8c
- hrx: support split glu binary variants
  rsuderman
  committed
  c196154
- hrx: support large decode projections
  rsuderman
  committed
  e84f9cf
- hrx: elide zero-output graph nodes
  rsuderman
  committed
  49a0f07
- hrx: restrict flash attention to 128-wide heads
  rsuderman
  committed
  4bfe977
- hrx: support 256-wide prefill flash attention
  rsuderman
  committed
  701e725
- hrx: support wider prefill flash attention heads
  rsuderman
  committed
  e5c2f9a
- hrx: support block-64 flash attention heads
  rsuderman
  committed
  c93de09
- Fix Qwen MoE weighted reduce residual binding
  rsuderman
  committed
  89ef405
- Fix HRX Loom artifact and BF16 linkage
  rsuderman
  committed
  5890fad
Commits on Sep 5, 2026
- hrx: add generic gfx1151 graph optimizations
  jammm
  committed
  aadd602
Commits on Sep 7, 2026
- hrx: optimize Qwen3.8 decode on gfx1151
  jammm
  committed
  3f96e79
Commits on Sep 8, 2026
- Fix HRX graph view value publication
  rsuderman
  committed
  12b7037
- HRX: elide safe zero-byte compute ops
  rsuderman
  committed
  f70ce6f
- hrx: add common scale dispatch support
  rsuderman
  committed
  6193007
Commits on Sep 9, 2026
- Merge remote-tracking branch 'rocm-temp/users/jam/qwen38_27b_udq4kxl-v2-port' into hrx-loom-reorg-jam
  rsuderman
  committed
  ab89746
Commits on Sep 10, 2026
- Support HRX scale on safe F32 views
  rsuderman
  committed
  3bba794
Commits on Sep 11, 2026
- Support legacy quantized HRX matmul
  rsuderman
  committed
  0e4dc49
- Support strided F32 binary HRX inputs
  rsuderman
  committed
  829fee0
- Expand HRX GET_ROWS embedding coverage
  rsuderman
  committed
  2063bb4
- Expand HRX RoPE coverage
  rsuderman
  committed
  798060d
- Support HRX F32 dim0 concat with strided inputs
  rsuderman
  committed
  284d4cf
- Fix HRX RoPE span bounds for strided inputs
  rsuderman
  committed
  84411f4
- Expand HRX quantized matmul coverage
  rsuderman
  committed
  a6687fb
- Support strided HRX SET_ROWS cache updates
  rsuderman
  committed
  28ed1c7
- Support strided HRX RMS norm inputs
  rsuderman
  committed
  b5d5d1e
- Support HRX F32 CONT materialization
  rsuderman
  committed
  94f76ba
- hrx: support asymmetric flash attention heads
  rsuderman
  committed
  6469564
- Support packed GLU dispatches
  rsuderman
  committed
  ac39e9c
Commits on Sep 12, 2026
- Add generic HRX SSM_CONV fallback
  rsuderman
  committed
  07fb324
- Support hidden-size 64 HRX RMS norm
  rsuderman
  committed
  f1dc43a
- Disable HRX symmetric matmul routes
  rsuderman
  committed
  0fde9e4
- Add decode QKV attention fusion kernels
  rsuderman
  committed
  853df8b
- Add decode matmul fusion coverage
  rsuderman
  committed
  27f9c3f
- Refactor SwiGLU decode onto shared dual matmul core
  rsuderman
  committed
  a0458ba
Commits on Sep 13, 2026
- Generalize Q6_K packed skinny matmul routing
  rsuderman
  committed
  529fa0b
Commits on Sep 15, 2026
- Use stream buffer copies for HRX host staging
  rsuderman
  committed
  ff62cfa
- Add serial debug mode for HRX command programs
  rsuderman
  committed
  b62031d
- hrx : optimize dense Qwen3.8 prefill and MTP decode
  
  jammm
  authored and
  rsuderman
  committed
  6c4a83f
- hrx: fix loom cherry-pick compatibility
  rsuderman
  committed
  c7c840e
- hrx: support updated Loom artifact API
  rsuderman
  committed
  ce23d99

Commits

Commits on Sep 3, 2026

Commits on Sep 5, 2026

Commits on Sep 7, 2026

Commits on Sep 8, 2026

Commits on Sep 9, 2026

Commits on Sep 10, 2026

Commits on Sep 11, 2026

Commits on Sep 12, 2026

Commits on Sep 13, 2026

Commits on Sep 15, 2026