CUDA: small PP performance improvement for MoE models (#589)

ikawrakow · Iwan Kawrakow · web-flow · commit 4c0b66026619 · 2025-07-07T07:23:12.000+02:00
* Trying to implement quantized fmoe - not working yet

* This works, but is slower than the non-working version

* quantize_mmq_q8_1_id

* Minor

---------

Co-authored-by: Iwan Kawrakow &lt;iwan.kawrakow@gmail.com&gt;
diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
@@ -2186,6 +2186,7 @@ struct mmid_row_mapping {
     int32_t i2;
 };
 
+template <typename data_t = float>
 static __global__ void k_copy_src_to_contiguous(const char * __restrict__ src_original, char * __restrict__ src_contiguous,
                                                   const mmid_row_mapping * __restrict__ row_mapping,
                                                   int64_t ne10, int64_t ne11, size_t nb11, size_t nb12) {
@@ -2194,8 +2195,8 @@ static __global__ void k_copy_src_to_contiguous(const char * __restrict__ src_or
     const int32_t i11 = row_mapping[i].i1 % ne11;
     const int32_t i12 = row_mapping[i].i2;
 
-    float * src_row_contiguous = (float *)(src_contiguous + i*nb11);
-    const float * src_row_original = (const float *)(src_original + i11*nb11 + i12*nb12);
+    data_t * src_row_contiguous = (data_t *)(src_contiguous + i*nb11);
+    const data_t * src_row_original = (const data_t *)(src_original + i11*nb11 + i12*nb12);
 
     for (int j = threadIdx.x; j < ne10; j += blockDim.x) {
         src_row_contiguous[j] = src_row_original[j];
@@ -2673,6 +2674,17 @@ static bool ggml_cuda_up_gate_unary(ggml_backend_cuda_context & ctx, ggml_tensor
             }
         }
     } else {
+        //printf("ne10 = %ld, ne11 = %ld, ne12 = %ld, nb10 = %zu nb11 = %zu nb12 = %zu\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->nb[0], src1->nb[1], src1->nb[2]);
+        ggml_cuda_pool_alloc<char> src1_quantized(ctx.pool());
+        bool use_quantized_src1 = false;
+        int64_t src1_padded_num_cols = 0, src1_padded_row_size = 0, src1_quantized_size = 0;
+        if (ggml_is_quantized(src0_1->type) && src0_1->type == src0_2->type && src1->ne[1] == 1 && src1->ne[3] == 1) {
+            src1_padded_num_cols = GGML_PAD(src1->ne[0], MATRIX_ROW_PADDING);
+            src1_padded_row_size = src1_padded_num_cols/ggml_blck_size(GGML_TYPE_Q8_1)*ggml_type_size(GGML_TYPE_Q8_1);
+            src1_quantized_size  = src1_padded_row_size*src1->ne[2] + get_mmq_x_max_host(ggml_cuda_info().devices[ctx.device].cc)*sizeof(block_q8_1_mmq);
+            src1_quantized.alloc(src1_quantized_size);
+            use_quantized_src1 = true;
+        }
         ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
         ggml_cuda_pool_alloc<char>  dst_up_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
         ggml_cuda_pool_alloc<char>  dst_gate_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
@@ -2704,7 +2716,13 @@ static bool ggml_cuda_up_gate_unary(ggml_backend_cuda_context & ctx, ggml_tensor
             if (num_src1_rows == 0) continue;
             size_t mapping_offset = cum_moe_counts[i02];
 
-            {
+            if (use_quantized_src1) {
+                quantize_mmq_q8_1_id_cuda((const float *)src1->data, src1_quantized.get(), (const char *)(dev_row_mapping.get() + mapping_offset),
+                        src1->ne[0], num_src1_rows, src1_padded_num_cols, src0_1->type, stream);
+                CUDA_CHECK(cudaGetLastError());
+                src1_row.data = src1_quantized.get();
+            }
+            else {
                 dim3 block_dims(std::min((unsigned int)ne10, 768u));
                 dim3 grid_dims(num_src1_rows);
                 k_copy_src_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
@@ -2719,21 +2737,31 @@ static bool ggml_cuda_up_gate_unary(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(nb1 == sizeof(float)*ne0);
 
             src1_row.ne[1] = num_src1_rows;
-            src1_row.nb[1] = nb11;
-            src1_row.nb[2] = num_src1_rows*nb11;
-            src1_row.nb[3] = num_src1_rows*nb11;
+            src1_row.nb[1] = use_quantized_src1 ? src1_padded_row_size : nb11;
+            src1_row.nb[2] = num_src1_rows*src1_row.nb[1];
+            src1_row.nb[3] = num_src1_rows*src1_row.nb[1];
 
             dst_row.ne[1] = num_src1_rows;
             dst_row.nb[1] = nb1;
             dst_row.nb[2] = num_src1_rows*nb1;
             dst_row.nb[3] = num_src1_rows*nb1;
 
             dst_row.data  =  dst_up_contiguous.get();
-            ggml_cuda_mul_mat(ctx, &src0_1_row, &src1_row, &dst_row);
+            if (use_quantized_src1) {
+                ggml_cuda_op_mul_mat_q(ctx, &src0_1_row, &src1_row, &dst_row, (const char *)src0_1_row.data, nullptr, src1_quantized.get(), (float *)dst_row.data,
+                        0, src0_1_row.ne[1], num_src1_rows, src1_padded_num_cols, stream);
+            } else {
+                ggml_cuda_mul_mat(ctx, &src0_1_row, &src1_row, &dst_row);
+            }
             CUDA_CHECK(cudaGetLastError());
 
             dst_row.data  = dst_gate_contiguous.get();
-            ggml_cuda_mul_mat(ctx, &src0_2_row, &src1_row, &dst_row);
+            if (use_quantized_src1) {
+                ggml_cuda_op_mul_mat_q(ctx, &src0_2_row, &src1_row, &dst_row, (const char *)src0_2_row.data, nullptr, src1_quantized.get(), (float *)dst_row.data,
+                        0, src0_2_row.ne[1], num_src1_rows, src1_padded_num_cols, stream);
+            } else {
+                ggml_cuda_mul_mat(ctx, &src0_2_row, &src1_row, &dst_row);
+            }
             CUDA_CHECK(cudaGetLastError());
 
             ggml_fused_mul_unary(ctx, (ggml_unary_op)dst->op_params[0], ggml_nelements(&dst_row),
diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu
@@ -166,6 +166,98 @@ static __global__ void quantize_mmq_q8_1(
     }
 }
 
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+template <mmq_q8_1_ds_layout ds_layout>
+static __global__ void quantize_mmq_q8_1_id(
+    const float * __restrict__ x, void * __restrict__ vy, const char * row_mapping, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+
+    constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
+    constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
+
+    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+
+    if (ix0 >= kx0_padded) {
+        return;
+    }
+
+    const float4 * x4 = (const float4 *) x;
+
+    const mmid_row_mapping * mapping = (const mmid_row_mapping *)row_mapping;
+    const int64_t ii = mapping[blockIdx.y].i2;
+
+    block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
+
+    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
+    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
+    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+
+    // Load 4 floats per thread and calculate max. abs. value between them:
+    const float4 xi = ix0 < kx0 ? x4[(ii*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    float amax = fabsf(xi.x);
+    amax = fmaxf(amax, fabsf(xi.y));
+    amax = fmaxf(amax, fabsf(xi.z));
+    amax = fmaxf(amax, fabsf(xi.w));
+
+    // Exchange max. abs. value between vals_per_scale/4 threads.
+#pragma unroll
+    for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+    }
+
+    float sum;
+    if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
+        sum = xi.x + xi.y + xi.z + xi.w;
+
+        // Exchange calculate sum across vals_per_sum/4 threads.
+#pragma unroll
+        for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) {
+            sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE);
+        }
+    }
+
+    const float d = amax/127.f;
+    const float d_inv = d > 0 ? 1/d : 0.f;
+    char4 q;
+    q.x = roundf(xi.x*d_inv);
+    q.y = roundf(xi.y*d_inv);
+    q.z = roundf(xi.z*d_inv);
+    q.w = roundf(xi.w*d_inv);
+
+    // Write back 4 int8 values as a single 32 bit value for better memroy bandwidth:
+    char4 * yqs4 = (char4 *) y[ib].qs;
+    yqs4[iqs/4] = q;
+
+    if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6) {
+        if (iqs % 16 != 0 || iqs >= 96) {
+            return;
+        }
+
+        y[ib].d2s6[2 + iqs/16] = sum;
+
+        if (iqs % 64 != 0) {
+            return;
+        }
+
+        y[ib].d2s6[iqs/64] = d;
+
+        return;
+    }
+
+    if (iqs % 32 != 0) {
+        return;
+    }
+
+    if (ds_layout == MMQ_Q8_1_DS_LAYOUT_DS4) {
+        y[ib].ds4[iqs/32] = make_half2(d, sum);
+    } else {
+        y[ib].d4[iqs/32]  = d;
+    }
+}
+
 void quantize_row_q8_1_cuda(
     const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
     const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
@@ -208,6 +300,35 @@ void quantize_mmq_q8_1_cuda(
     }
 }
 
+void quantize_mmq_q8_1_id_cuda(
+    const float * x, void * vy, const char * row_mapping, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded,
+    const ggml_type type_x, cudaStream_t stream) {
+
+    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
+
+    const int64_t block_num_x = (kx0_padded + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(block_num_x, kx1, 1);
+    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
+    switch (mmq_get_q8_1_ds_layout(type_x)) {
+        case MMQ_Q8_1_DS_LAYOUT_D4:
+            quantize_mmq_q8_1_id<MMQ_Q8_1_DS_LAYOUT_D4>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, row_mapping, kx0, kx1, kx0_padded);
+            break;
+        case MMQ_Q8_1_DS_LAYOUT_DS4:
+            quantize_mmq_q8_1_id<MMQ_Q8_1_DS_LAYOUT_DS4>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, row_mapping, kx0, kx1, kx0_padded);
+            break;
+        case MMQ_Q8_1_DS_LAYOUT_D2S6:
+            quantize_mmq_q8_1_id<MMQ_Q8_1_DS_LAYOUT_D2S6>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, row_mapping, kx0, kx1, kx0_padded);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+
 void quantize_tensor_q8_1_cuda(const struct ggml_tensor * src, void * vy, const enum ggml_type type, cudaStream_t stream) {
     GGML_ASSERT(src->ne[1] == 1 && src->ne[3] == 1);
     GGML_ASSERT(src->type == GGML_TYPE_F32);
diff --git a/ggml/src/ggml-cuda/quantize.cuh b/ggml/src/ggml-cuda/quantize.cuh
@@ -30,5 +30,9 @@ void quantize_mmq_q8_1_cuda(
     const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
     const ggml_type type_x, cudaStream_t stream);
 
+void quantize_mmq_q8_1_id_cuda(
+    const float * x, void * vy, const char * row_mapping, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded,
+    const ggml_type type_x, cudaStream_t stream);
+
 // For now only applicable for tensors with ne[1] = 1, ne[3] = 1, and useful if ne[2] > 1
 void quantize_tensor_q8_1_cuda(const struct ggml_tensor * src, void * vy, const enum ggml_type type, cudaStream_t stream);
