CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl: Add reorder to Q6_K mmvq implementation (#13885)

Browse Source 
* Add Reorder to Q6_K mmvq implementation

* Address PR comments: clean up comments

* Remove unused parameter after refactoring q4_k

* Adding inline to function and removing unnecessary reference to int

---------

Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
This commit is contained in:
Nicolò Scipione
2025-06-09 11:47:07 +02:00
committed by GitHub
parent 91a8ee6a6f
commit b460d16ae8
6 changed files with 244 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
ggml/src/ggml-sycl/convert.cpp
View File

@@ -265,6 +265,17 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
#endif
}
template <typename dst_t>
static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
const int64_t nb = k / QK_K;
dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
stream->parallel_for(
sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
[=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
}
template <typename dst_t>
static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
dpct::queue_ptr stream) {
@@ -530,7 +541,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
case GGML_TYPE_Q5_K:
return dequantize_row_q5_K_sycl;
case GGML_TYPE_Q6_K:
return dequantize_row_q6_K_sycl;
if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
return dequantize_row_q6_K_sycl_reorder;
} else {
return dequantize_row_q6_K_sycl;
}
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_sycl;
case GGML_TYPE_IQ1_M:
@@ -587,7 +602,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
case GGML_TYPE_Q5_K:
return dequantize_row_q5_K_sycl;
case GGML_TYPE_Q6_K:
return dequantize_row_q6_K_sycl;
if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
return dequantize_row_q6_K_sycl_reorder;
} else {
return dequantize_row_q6_K_sycl;
}
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_sycl;
case GGML_TYPE_IQ1_M: