LLM inference in C/C++
C++ ggml
webgpu : add Q2_0 support (#30914)
Chanho Lee0b1281aeef18
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp b/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hppindex 6efff9b376..e01a417704 100644--- a/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp+++ b/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp@@ -1655,6 +1655,7 @@ class ggml_webgpu_shader_lib {switch (key.src_type) {case GGML_TYPE_Q1_0:+ case GGML_TYPE_Q2_0:case GGML_TYPE_Q4_0:case GGML_TYPE_Q5_0:case GGML_TYPE_Q8_0:@@ -1699,7 +1700,7 @@ class ggml_webgpu_shader_lib {} else if ((key.src_type >= GGML_TYPE_Q4_0 && key.src_type <= GGML_TYPE_Q8_1) ||key.src_type == GGML_TYPE_IQ4_NL || key.src_type == GGML_TYPE_MXFP4) {defines.push_back("BLOCK_SIZE=32u");- } else if (key.src_type == GGML_TYPE_NVFP4) {+ } else if (key.src_type == GGML_TYPE_Q2_0 || key.src_type == GGML_TYPE_NVFP4) {defines.push_back("BLOCK_SIZE=64u");} else if (key.src_type >= GGML_TYPE_Q2_K) {defines.push_back("BLOCK_SIZE=256u");diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cppindex 77fbb4f476..71d9ae2b92 100644--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp@@ -4395,6 +4395,7 @@ static bool ggml_backend_webgpu_device_supports_buft(ggml_backend_dev_t dev, ggmstatic bool ggml_webgpu_supported_qtype(ggml_type type) {switch (type) {case GGML_TYPE_Q1_0:+ case GGML_TYPE_Q2_0:case GGML_TYPE_Q4_0:case GGML_TYPE_Q4_1:case GGML_TYPE_Q5_0:@@ -4500,6 +4501,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, constcase GGML_TYPE_F16:case GGML_TYPE_BF16:case GGML_TYPE_Q1_0:+ case GGML_TYPE_Q2_0:case GGML_TYPE_Q4_0:case GGML_TYPE_Q4_1:case GGML_TYPE_Q5_0:diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgslindex a3114a4bea..4bf3c93471 100644--- a/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl+++ b/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl@@ -51,6 +51,20 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {}#endif+#ifdef Q2_0+fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {+ let block_byte_base = (src_base + offset) * 18u;+ let d = load_f16_as_f32_at_src(block_byte_base);+ for (var j = 0u; j < 4u; j++) {+ let q_packed = load_u32_at_src(block_byte_base + 2u + j * 4u);+ for (var k = 0u; k < 16u; k++) {+ let q = (q_packed >> (2u * k)) & 3u;+ dst[dst_base + offset * 64u + j * 16u + k] = (f32(q) - 1.0) * d;+ }+ }+}+#endif+#ifdef Q4_0fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {let block_byte_base = (src_base + offset) * 18; // Block stride: 18 bytesdiff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmplindex 83f6fa1728..1af6286ac4 100644--- a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl@@ -111,6 +111,29 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3}#endif // INIT_SRC0_SHMEM_Q1_0+#ifdef INIT_SRC0_SHMEM_Q2_0+const BLOCK_SIZE = 64u;+const BLOCK_SIZE_BYTES = 18u;+const NQ = 16u;++fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {+ for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {+ let global_m = offset_m + i / TILE_K;+ let global_k = k_outer + i % TILE_K;+ if (global_m < params.m && global_k < params.k) {+ let block_idx = batch_offset + global_m * params.stride_01 + global_k / BLOCK_SIZE;+ let block_byte_base = block_idx * BLOCK_SIZE_BYTES;+ let d = load_f16_at_src0(block_byte_base);+ let q_packed = load_u32_at_src0(block_byte_base + 2u + (global_k % BLOCK_SIZE) / 4u);+ for (var k = 0u; k < NQ; k++) {+ let q = (q_packed >> (2u * k)) & 3u;+ shmem[i + k] = (f16(q) - 1.0) * d;+ }+ }+ }+}+#endif // INIT_SRC0_SHMEM_Q2_0+// legacy-quants#if defined(INIT_SRC0_SHMEM_Q4_0) || defined(INIT_SRC0_SHMEM_Q4_1) || defined(INIT_SRC0_SHMEM_Q5_0) || defined(INIT_SRC0_SHMEM_Q5_1) || defined(INIT_SRC0_SHMEM_Q8_0) || defined(INIT_SRC0_SHMEM_Q8_1) || defined(INIT_SRC0_SHMEM_MXFP4)const BLOCK_SIZE = 32u;diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmplindex 841777df9c..2c116aad6c 100644--- a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl@@ -119,6 +119,46 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src}#endif+#ifdef MUL_ACC_Q2_0+#define BLOCK_SIZE 64+#define BLOCK_SIZE_BYTES 18+#define THREADS_PER_BLOCK 4+#define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)+fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;++ let num_blocks = params.k / BLOCK_SIZE;+ let thread_within_block = thread_id % THREADS_PER_BLOCK;+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * ELEMS_PER_THREAD;+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;+ for (var col = 0u; col < NUM_COLS; col++) {+ for (var i = 0u; i < ELEMS_PER_THREAD; i++) {+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);+ }+ }+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {+ let output_row = row_base + row;+ if (output_row < params.m) {+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;+ let d = f32(load_f16_at_src0(block_byte_base));+ let q_packed = load_u32_at_src0(block_byte_base + 2u + 4u * thread_within_block);+ for (var col = 0u; col < NUM_COLS; col++) {+ var row_sum = 0.0;+ for (var i = 0u; i < ELEMS_PER_THREAD; i++) {+ let q = (q_packed >> (2u * i)) & 3u;+ row_sum += (f32(q) - 1.0) * x_block[col][i];+ }+ acc[col][row] += d * row_sum;+ }+ }+ }+ }++ return acc;+}+#endif+#ifdef MUL_ACC_Q4_0#define BLOCK_SIZE 32#define BLOCK_SIZE_BYTES 18