ROCm: validate native gfx1151 serving on Strix Halo - #260

Open

dbourdea wants to merge 21 commits into

FlashML-org:mainfrom

dbourdea:amd-rocm-gfx1151

Commits on Aug 28, 2026

rocm: fix native extensions, kernel JIT builds, and Triton PTX fallbacks for gfx1150

skywalk1411
authored and
David
committed
809e0ee
rocm: fix fp8 native-capability detection and the offload cache's fast-copy kernel

skywalk1411
authored and
David
committed
7d1a033
rocm: fix the GGUF kernel JIT build (kernel/gguf.py, dispatch.h)

skywalk1411
authored and
David
committed
dd2fc2e
rocm: revert two defensive fixes that turned out to be unnecessary

skywalk1411
authored and
David
committed
fea3f6b
feat(rocm): harden HIP runtime gating for gfx1151
David
committed
a8a650c
fix(rocm): accept HIP tensors in TVM JIT kernels
David
committed
5db6157
fix(rocm): support HIP fast-index copy tensors
David
committed
ebac382
fix(rocm): skip unsafe optional NVFP4 prefill warmup
David
committed
f0d4d12
fix(rocm): use safe Triton NVFP4 prefill path
David
committed
df8bc43
fix(rocm): locate Thrust headers for GGUF JIT
David
committed
514af56
fix(rocm): pass GGUF Thrust headers as system include
David
committed
2042dda
fix(rocm): locate HIP runtime for GGUF JIT
David
committed
065d806
docs(rocm): record LAN-223 full model validation
David
committed
81592bb
docs(rocm): retain the GGUF HIP extension cache
David
committed
b589a53
docs(rocm): add warm LAN-223 TPS and GGUF cache reuse
David
committed
c3cc2b7
docs(rocm): compare Gemma HIP throughput with llama.cpp Vulkan
David
committed
3724569
docs(rocm): compare FreeToken with llama.cpp ROCm 10
David
committed
b5005ee
perf(rocm): target GGUF HIP extensions to active gfx
David
committed
cd2f10f
perf(rocm): retain conservative GGUF math flags
David
committed
9a89566
docs(rocm): record LAN-223 TPS optimization results
David
committed
16fe07e
docs(rocm): record profiler limitations
David
committed
0977c80

Navigation Menu

ROCm: validate native gfx1151 serving on Strix Halo - #260

ROCm: validate native gfx1151 serving on Strix Halo#260
dbourdea wants to merge 21 commits into
FlashML-org:mainfrom
dbourdea:amd-rocm-gfx1151

Commits on Aug 28, 2026

rocm: fix native extensions, kernel JIT builds, and Triton PTX fallbacks for gfx1150

rocm: fix fp8 native-capability detection and the offload cache's fast-copy kernel

rocm: fix the GGUF kernel JIT build (kernel/gguf.py, dispatch.h)

rocm: revert two defensive fixes that turned out to be unnecessary

feat(rocm): harden HIP runtime gating for gfx1151

fix(rocm): accept HIP tensors in TVM JIT kernels

fix(rocm): support HIP fast-index copy tensors

fix(rocm): skip unsafe optional NVFP4 prefill warmup

fix(rocm): use safe Triton NVFP4 prefill path

fix(rocm): locate Thrust headers for GGUF JIT

fix(rocm): pass GGUF Thrust headers as system include

fix(rocm): locate HIP runtime for GGUF JIT

docs(rocm): record LAN-223 full model validation

docs(rocm): retain the GGUF HIP extension cache

docs(rocm): add warm LAN-223 TPS and GGUF cache reuse

docs(rocm): compare Gemma HIP throughput with llama.cpp Vulkan

docs(rocm): compare FreeToken with llama.cpp ROCm 10

perf(rocm): target GGUF HIP extensions to active gfx

perf(rocm): retain conservative GGUF math flags

docs(rocm): record LAN-223 TPS optimization results

docs(rocm): record profiler limitations