Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

sycl: add Q2_K to DMMV reorder path (#25064)

Browse Source 
Signed-off-by: Todd Malsbary <todd.malsbary@intel.com>
This commit is contained in:
Todd Malsbary
2026-07-13 09:53:39 +03:00
committed by GitHub
parent 38fd5c9993
commit e474bba7af
2 changed files with 168 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/ggml-sycl.cpp
+48
View File
@@ -547,6 +547,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
switch (tensor->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
@@ -3675,6 +3676,7 @@ inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
return true;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
@@ -3690,6 +3692,7 @@ inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) {
case GGML_TYPE_Q1_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q5_K:
@@ -4069,6 +4072,49 @@ static bool reorder_qw_q6_k_moe(uint8_t * data_device, size_t expert_bytes, int6
return true;
}
static bool reorder_qw_q2_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
GGML_ASSERT(size % sizeof(block_q2_K) == 0);
GGML_ASSERT(offset % sizeof(block_q2_K) == 0);
const int nblocks = size / sizeof(block_q2_K);
sycl_reorder_temp_buffer tmp(stream, size);
if (!tmp) {
GGML_LOG_WARN("%s: failed to allocate %zu bytes for reorder temp buffer, skipping reorder\n", __func__, size);
return false;
}
uint8_t * tmp_buf = static_cast<uint8_t *>(tmp.ptr);
sycl::event copy_event;
SYCL_CHECK(CHECK_TRY_ERROR(copy_event = stream->memcpy(tmp_buf, data_device, size)));
if (!g_ggml_sycl_use_async_mem_op) {
copy_event.wait();
}
auto * qs_ptr = data_device;
auto * scales_ptr = qs_ptr + (QK_K / 4) * nblocks;
sycl::half2 * dm_ptr = (sycl::half2 *) (scales_ptr + (QK_K / 16) * nblocks);
auto reorder_event = stream->parallel_for(nblocks, [=](auto i) {
const block_q2_K * x = (const block_q2_K *) tmp_buf;
const int ib = i;
for (int j = 0; j < QK_K / 4; ++j) {
qs_ptr[ib * (QK_K / 4) + j] = x[ib].qs[j];
}
for (int j = 0; j < QK_K / 16; ++j) {
scales_ptr[ib * (QK_K / 16) + j] = x[ib].scales[j];
}
dm_ptr[ib] = x[ib].dm;
});
if (!g_ggml_sycl_use_async_mem_op) {
reorder_event.wait_and_throw();
}
return true;
}
static bool reorder_qw_q3_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
GGML_ASSERT(size % sizeof(block_q3_K) == 0);
GGML_ASSERT(offset % sizeof(block_q3_K) == 0);
@@ -4245,6 +4291,8 @@ static bool reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
return reorder_qw_q4_0(data_device, ncols, nrows, size, 0, stream);
case GGML_TYPE_Q8_0:
return reorder_qw_q8_0(data_device, ncols, nrows, size, 0, stream);
case GGML_TYPE_Q2_K:
return reorder_qw_q2_k(data_device, size, 0, stream);
case GGML_TYPE_Q3_K:
return reorder_qw_q3_k(data_device, size, 0, stream);
case GGML_TYPE_Q4_K: