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

ggml-webgpu: add mulmat with overlapping src0/src1 (e.g., for minimax-01) (#27321)

Browse Source
This commit is contained in:
Masashi Yoshimura
2026-08-19 16:29:33 +09:00
committed by GitHub
parent 0adcc3bb57
commit 5112b9738b
10 changed files with 189 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+40 -23
View File
@@ -1628,48 +1628,65 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
// Get or create pipeline
webgpu_pipeline pipeline;
std::vector<webgpu_dispatch_desc> dispatches;
const bool src_overlap = ggml_webgpu_tensor_binding_overlap(ctx->global_ctx, src0, src1) && !use_mmvq;
if (use_mat_vec) {
if (use_mmvq) {
ggml_webgpu_quantize_q8_dispatch(ctx, src0, src1, dst, dispatches);
}
pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx);
pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx, src_overlap);
} else {
pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx);
pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx, src_overlap);
}
uint32_t offset_src0 = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type));
uint32_t offset_src1 = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type));
size_t merged_offset = 0;
size_t merged_size = 0;
if (src_overlap) {
const ggml_webgpu_merged_binding_range merged_range =
ggml_webgpu_tensor_merged_binding_range(ctx, { src0, src1 });
merged_offset = merged_range.offset;
merged_size = merged_range.size;
offset_src0 = ggml_webgpu_tensor_merged_element_offset(src0, merged_range);
offset_src1 = ggml_webgpu_tensor_merged_element_offset(src1, merged_range);
}
// Build params
std::vector<uint32_t> params = {
(uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
(uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
(uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
(uint32_t) dst->ne[0],
(uint32_t) dst->ne[1],
(uint32_t) src0->ne[0],
(uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
(uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
(uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
(uint32_t) src0->ne[2],
(uint32_t) src0->ne[3],
(uint32_t) (src1->ne[2] / src0->ne[2]),
(uint32_t) (src1->ne[3] / src0->ne[3])
};
std::vector<uint32_t> params = { offset_src0,
offset_src1,
(uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
(uint32_t) dst->ne[0],
(uint32_t) dst->ne[1],
(uint32_t) src0->ne[0],
(uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
(uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
(uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
(uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
(uint32_t) src0->ne[2],
(uint32_t) src0->ne[3],
(uint32_t) (src1->ne[2] / src0->ne[2]),
(uint32_t) (src1->ne[3] / src0->ne[3]) };
// Build bind group entries
std::vector<wgpu::BindGroupEntry> entries = {};
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
if (use_mmvq) {
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
auto & mmvq_qq8_entry = dispatches[0].bind_group_entries[1];
entries.push_back(ggml_webgpu_make_bind_group_entry(1, ggml_webgpu_tensor_buf(dst), mmvq_qq8_entry.offset,
mmvq_qq8_entry.size));
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
} else if (src_overlap) {
entries.push_back(
ggml_webgpu_make_bind_group_entry(0, ggml_webgpu_tensor_buf(src0), merged_offset, merged_size));
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
} else {
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1));
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
}
entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
// Calculate workgroup dimensions
uint32_t wg_x = 1;