happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml-cuda : fix CDNA2 compute capability constant for gfx90a (MI210) (#21519) 

GGML_CUDA_CC_CDNA2 was set to 0x910
Fix by setting the constant to 0x90a to match the actual gfx90a ISA.
This commit is contained in:
Antoine Viallon 2026-04-07 12:18:55 +02:00 committed by GitHub
parent ecce0087da
commit 71a81f6fcc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/common.cuh
View File

@ -65,7 +65,7 @@
#define GGML_CUDA_CC_VEGA (GGML_CUDA_CC_OFFSET_AMD + 0x900) // Vega56/64, minimum for fp16 dual issue
#define GGML_CUDA_CC_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 0x906) // MI50/Radeon VII, minimum for dp4a
#define GGML_CUDA_CC_CDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x908) // MI100, minimum for MFMA, acc registers
#define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x910) // MI210, minimum acc register renameing
#define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x90a) // MI210 (gfx90a), minimum acc register renaming
#define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x942) // MI300
// RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32