ggml-webgpu: add support for NVFP4 (#25143)

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp b/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
index c00a2e9ee9b86cf9cb690697e30d72a3b57bb6a2..80a16d16d23662666943fc4baa3b26bf1cff0d73 100644 (file)
--- a/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp
@@ -1563,6 +1563,7 @@ class ggml_webgpu_shader_lib {
                         case GGML_TYPE_IQ1_S:
                         case GGML_TYPE_IQ4_NL:
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             {
                                 // Quantized types using u32 buffers for portability.
                                 defines.push_back("SRC_TYPE=u32");
@@ -1593,6 +1594,8 @@ 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) {
+                        defines.push_back("BLOCK_SIZE=64u");
                     } else if (key.src_type >= GGML_TYPE_Q2_K) {
                         defines.push_back("BLOCK_SIZE=256u");
                     } else {
@@ -1960,6 +1963,7 @@ class ggml_webgpu_shader_lib {
                             defines.push_back(type_upper + "_TABLES");
                             break;
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             defines.push_back(type_upper + "_LUT");
                             break;
                         default:
@@ -2103,6 +2107,7 @@ class ggml_webgpu_shader_lib {
                             defines.push_back(type_upper + "_TABLES");
                             break;
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             defines.push_back(type_upper + "_LUT");
                             break;
                         default:
@@ -2274,6 +2279,7 @@ class ggml_webgpu_shader_lib {
                             defines.push_back(type_upper + "_TABLES");
                             break;
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             defines.push_back(type_upper + "_LUT");
                             break;
                         default:
@@ -2394,6 +2400,7 @@ class ggml_webgpu_shader_lib {
                             defines.push_back(type_upper + "_TABLES");
                             break;
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             defines.push_back(type_upper + "_LUT");
                             break;
                         default:

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index f0ec18abd9afb263f337e3609fbdd08f29ac2147..29025e9ba4e391f0ee21376d40a306ed8cf10c8f 100644 (file)
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -4056,6 +4056,7 @@ static bool ggml_webgpu_supported_qtype(ggml_type type) {
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
             return true;
         default:
             return false;
@@ -4156,6 +4157,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                             case GGML_TYPE_IQ4_NL:
                             case GGML_TYPE_IQ4_XS:
                             case GGML_TYPE_MXFP4:
+                            case GGML_TYPE_NVFP4:
                                 supports_op = true;
                                 break;
                             default:
@@ -4196,6 +4198,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                         case GGML_TYPE_IQ4_NL:
                         case GGML_TYPE_IQ4_XS:
                         case GGML_TYPE_MXFP4:
+                        case GGML_TYPE_NVFP4:
                             supports_op = true;
                             break;
                         default:

diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl b/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl
index 758efa17d779c74ead15864ad59f9a697a2f35cc..6634fbd65782110b3c66ba8e7a043a6a07bc95b8 100644 (file)
--- a/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl
@@ -896,9 +896,23 @@ const kvalues_iq4nl = array<i32, 16>(
 
 #endif
 
-#ifdef MXFP4_LUT
+#if defined(MXFP4_LUT) || defined(NVFP4_LUT)
 const kvalues_mxfp4 = array<i32, 16>(
     0, 1, 2, 3, 4, 6, 8, 12, 0, -1, -2, -3, -4, -6, -8, -12
 );
-#endif
+#endif // MXFP4_LUT || NVFP4_LUT
 
+#ifdef NVFP4_LUT
+fn ue4m3_to_fp32(u: u32) -> f32 {
+    if (u == 0u || u == 127u) {
+        return 0.0;
+    }
+    let exp = (u >> 3u) & 15u;
+    let man = u & 7u;
+    if (exp == 0u) {
+        return f32(man) * (1.0 / 512.0);
+    }
+    let bits = ((exp + 120u) << 23u) | (man << 20u);
+    return bitcast<f32>(bits);
+}
+#endif // NVFP4_LUT

diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl
index 78d61a93d28cc389934e563085b46b81abf90a19..487edb3275f3080a88ab24fef5e7783cd7d72bea 100644 (file)
--- a/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl
@@ -672,6 +672,27 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
 }
 #endif
 
+#ifdef NVFP4
+fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
+    let block_byte_base = (src_base + offset) * 36;
+    let d_word = load_u32_at_src(block_byte_base);
+    for (var sub: u32 = 0u; sub < 4; sub++) {
+        let d = ue4m3_to_fp32(get_byte(d_word, sub)) * 0.5;
+        for (var j: u32 = 0u; j < 2; j++) {
+            let q_packed = load_u32_at_src(block_byte_base + 4 + sub * 8 + j * 4);
+            for (var k: u32 = 0; k < 4; k++) {
+                let q_byte = get_byte(q_packed, k);
+                let q_lo = f32(kvalues_mxfp4[q_byte & 0xFu]) * d;
+                let q_hi = f32(kvalues_mxfp4[(q_byte >> 4) & 0xF]) * d;
+                let dst_offset = dst_base + offset * 64 + sub * 16 + j * 4 + k;
+                dst[dst_offset] = q_lo;
+                dst[dst_offset + 8u] = q_hi;
+            }
+        }
+    }
+}
+#endif
+
 
 @group(0) @binding(0)
 var<storage, read_write> src: array<SRC_TYPE>;

diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl
index 6a2eb8c824e1eb261dd0dc3648799d31fd6dbac8..13996ab5157541f1fa2add017bde1edcf5e2d747 100644 (file)
--- a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl
@@ -241,7 +241,7 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
 #endif // INIT_SRC0_SHMEM_Q8_1
 
 #if defined(INIT_SRC0_SHMEM_MXFP4)
-            let block_byte_base = src0_idx * 17u;
+            let block_byte_base = src0_idx * 17u; // BLOCK_SIZE_BYTES = 17u;
             let eu8 = get_byte(load_u32_at_src0_aligned(block_byte_base), block_byte_base & 3u);
             let e = ldexp(1.0, i32(eu8) - 128);
 
@@ -263,6 +263,47 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
 }
 #endif // legacy-quants
 
+#if defined(INIT_SRC0_SHMEM_NVFP4)
+const BLOCK_SIZE = 64u;
+const BLOCK_SIZE_BYTES = 36u;
+const SUB_BLOCK_SIZE = 16u; // elements sharing one UE4M3 scale
+const NQ = 16u;
+const BYTES_PER_THREAD = 8u;
+const BYTES_PER_INNER_LOOP = 4u;
+
+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 tile_m = i / TILE_K;
+        let tile_k_start = i % TILE_K;
+        let global_m = offset_m + tile_m;
+        let global_k_start = k_outer + tile_k_start;
+
+        if (global_m >= params.m) {
+            break;
+        }
+
+        let block_k  = global_k_start / BLOCK_SIZE;
+        let sub_block      = (global_k_start % BLOCK_SIZE) / SUB_BLOCK_SIZE;
+        let src0_idx = batch_offset + global_m * params.stride_01 + block_k;
+
+        let block_byte_base = src0_idx * BLOCK_SIZE_BYTES;
+        let d_byte_base     = block_byte_base;
+        let qs_byte_base    = block_byte_base + 4u;
+
+        let d = ue4m3_to_fp32(get_byte(load_u32_at_src0_aligned(d_byte_base), sub_block)) * 0.5;
+
+        for (var j = 0u; j < BYTES_PER_THREAD / BYTES_PER_INNER_LOOP; j++) {
+            let q_packed = load_u32_at_src0_aligned(qs_byte_base + sub_block * 8u + j * 4u);
+            for (var k = 0u; k < BYTES_PER_INNER_LOOP; k++) {
+                let q_byte = get_byte(q_packed, k);
+                shmem[i + j * BYTES_PER_INNER_LOOP + k]      = f16(f32(kvalues_mxfp4[q_byte & 0xF]) * d);
+                shmem[i + j * BYTES_PER_INNER_LOOP + k + 8u] = f16(f32(kvalues_mxfp4[(q_byte >> 4) & 0xF]) * d);
+            }
+        }
+    }
+}
+#endif // INIT_SRC0_SHMEM_NVFP4
+
 // k-quants
 #if defined(INIT_SRC0_SHMEM_Q2_K) || defined(INIT_SRC0_SHMEM_Q3_K) || defined(INIT_SRC0_SHMEM_Q4_K) || defined(INIT_SRC0_SHMEM_Q5_K) || defined(INIT_SRC0_SHMEM_Q6_K)
 const BLOCK_SIZE = 256u;

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.tmpl
index b0703fe9062d412bd0289698bd021a994c482175..8fd0d1907cfcaede078c836f030cd663d544569e 100644 (file)
--- a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl
@@ -1505,3 +1505,49 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
     return acc;
 }
 #endif
+
+#ifdef MUL_ACC_NVFP4
+#define BLOCK_SIZE 64
+#define BLOCK_SIZE_BYTES 36
+#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 sub = 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 + sub * ELEMS_PER_THREAD;
+        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
+        for (var col = 0u; col < NUM_COLS;col += 1) {
+            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
+                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 8] = f32(src1[x_base + col * params.stride_11 + i + 8]);
+            }
+        }
+        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 = ue4m3_to_fp32(get_byte(load_u32_at_src0_aligned(block_byte_base), sub)) * 0.5;
+                let q_w0 = load_u32_at_src0_aligned(block_byte_base + 4u + 8u * sub);
+                let q_w1 = load_u32_at_src0_aligned(block_byte_base + 8u + 8u * sub);
+                for (var col = 0u;col < NUM_COLS;col += 1) {
+                    var row_sum = 0.0;
+                    for (var l = 0u; l < 8u; l++) {
+                        let q_word = select(q_w0, q_w1, l >= 4u);
+                        let q_byte = get_byte(q_word, l % 4u);
+                        let q_lo = f32(kvalues_mxfp4[q_byte & 0xFu]) * d;
+                        let q_hi = f32(kvalues_mxfp4[(q_byte >> 4u) & 0xFu]) * d;
+                        row_sum += q_lo * x_block[col][l];
+                        row_sum += q_hi * x_block[col][l + 8u];
+                    }
+                    acc[col][row] += row_sum;
+                }
+            }
+        }
+    }
+
+    return acc;
+}
+#endif