Enable gfx1250 ROCm architecture (#46516)

Signed-off-by: Douglas Lehr <Doug.Lehr@amd.com>
Signed-off-by: jpvillam <juan.villamizar@amd.com>
Signed-off-by: Daniel <danichan@amd.com>
Signed-off-by: Jaden Mathias <jaden.mathias@amd.com>
Signed-off-by: jpvillam <jpvillam@amd.com>
Signed-off-by: Juan Villamizar <100237675+jpvillam-amd@users.noreply.github.com>
Signed-off-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Signed-off-by: Jaden Mathias <Jaden.Mathias@amd.com>
Co-authored-by: Douglas Lehr <Doug.Lehr@amd.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Daniel <danichan@amd.com>
Co-authored-by: Jaden Mathias <jaden.mathias@amd.com>
Co-authored-by: jpvillam <jpvillam@amd.com>
Co-authored-by: root <root@ctheliosr-1b114-f01-2.mnb.dcgpu>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Co-authored-by: dllehr <douglehr@amd.com>
Co-authored-by: Daniel Chan <Daniel.Chan@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Juan Villamizar
2026-07-31 01:41:12 +00:00
committed by GitHub
co-authored by Douglas Lehr Claude Opus 4.7 Daniel Jaden Mathias jpvillam root Douglas Lehr dllehr Daniel Chan Andreas Karatzas
parent 553fcb82d5
commit 4f1da84eb5
33 changed files with 1850 additions and 104 deletions
+5
View File
@@ -14,6 +14,10 @@
TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
// vLLM custom ops for rocm
// skinny_gemms.cu (LLMM1/wvSplitK/wvSplitKrc/wvSplitKQ) is excluded on gfx1250
// (gfx9/gfx11 ISA, unsupported there); skip these registrations to avoid
// undefined symbols. vLLM uses default/Triton GEMM for these ops on gfx1250.
#ifndef VLLM_SKIP_SKINNY_GEMMS
// Custom gemm op for matrix-vector multiplication
rocm_ops.def(
"LLMM1(Tensor in_a, Tensor in_b, int rows_per_block) -> "
@@ -46,6 +50,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
"Tensor scale_a, "
" Tensor scale_b, int CuCount) -> ()");
rocm_ops.impl("wvSplitKQ", torch::kCUDA, &wvSplitKQ);
#endif // VLLM_SKIP_SKINNY_GEMMS
#ifdef VLLM_ROCM_GFX1100
// W4A16 GPTQ kernels for AMD RDNA3 (gfx1100).