Skip to content

HIP/ROCm: two crash fixes for TurboQuant KV cache on RDNA #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Ooooze merged 2 commits into from
May 7, 2026
Merged

HIP/ROCm: two crash fixes for TurboQuant KV cache on RDNA #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2561d87
hip: add missing fattn-vec f16-turboN_0 template instances
May 7, 2026
425db5b
fattn-tile: fix GGML_ABORT for DKQ=DV=512 with gqa_ratio not div by 4
May 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 14 additions & 0 deletions ggml/src/ggml-cuda/fattn-tile.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,6 +68,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 64, 64)

GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
Expand Down Expand Up @@ -132,6 +133,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64)

GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 32, 64)
Expand Down Expand Up @@ -203,6 +205,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128)

GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
Expand Down Expand Up @@ -277,6 +280,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128)

GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4, 64, 64)
Expand Down Expand Up @@ -1259,6 +1263,16 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
return;
}

// DV > 256 (e.g. DKQ=DV=512, head_dim=512 models): extend GQA fallback to ncols2=2/1.
// Without this, gqa_ratio not divisible by 4 (e.g. ratio=2) reaches GGML_ABORT.
if (use_gqa_opt && gqa_ratio % 2 == 0) {
launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
return;
}

launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
return;
}
GGML_ABORT("fatal error");
}
Expand Down
3 changes: 3 additions & 0 deletions ggml/src/ggml-hip/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,14 +85,17 @@ else()
../ggml-cuda/template-instances/fattn-vec-instance-turbo3_0-turbo3_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo3_0-q8_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-q8_0-turbo3_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-f16-turbo3_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo2_0-turbo2_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo2_0-q8_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-q8_0-turbo2_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-f16-turbo2_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo3_0-turbo2_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo2_0-turbo3_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo4_0-turbo4_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo4_0-q8_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-q8_0-turbo4_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-f16-turbo4_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo4_0-turbo3_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo3_0-turbo4_0.cu
../ggml-cuda/template-instances/fattn-vec-instance-turbo4_0-turbo2_0.cu
Expand Down