MLA-2: Allow usage of q8_0 for KV cache on CUDA (#252)

* FlashMLA(CUDA): WIP to allow q8_0 quantized cache * WIP * FlashMLA(CUDA) - allow q8_0 for KV cache This works, and PP is not bad, but TG is still quite a bit slower. * FlashMLA(CUDA) - allow q8_0 for KV cache This is better. ~9% slower than f16 cache for short contexts, nearly on par at 16k tokens. --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2026-02-11 08:50:11 +00:00 · 2025-03-12 07:21:46 +02:00
parent 1266d12461
commit fc6a65dda4
2 changed files with 141 additions and 7 deletions
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -2296,9 +2296,6 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
                for (int64_t id = 0; id < n_ids; id++) {
                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);

-                    if (i02 < 0 || i02 >= n_as) continue;
-                    //GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
                    if (row_id_i != i02) {
                        continue;
                    }
@@ -3458,6 +3455,14 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
                    return true;
                }
+                if (ggml_is_contiguous(op->src[0]) && ggml_are_same_shape(op->src[0], op->src[1])) {
+                    if (src1_type == GGML_TYPE_F16 || src1_type == GGML_TYPE_BF16 || src1_type == GGML_TYPE_F32) {
+                        return true;
+                    }
+                }
+                if (ggml_are_same_shape(op->src[0], op->src[1]) && op->src[0]->type == GGML_TYPE_Q8_0 && op->src[1]->type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
                return false;
            } break;
        case GGML_OP_DUP:
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -1,4 +1,5 @@
 #include "cpy.cuh"
+#include "convert.cuh"

 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);

@@ -65,6 +66,66 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
    cpy_1(cx + x_offset, cdst + dst_offset);
 }

+//static __global__ void cpy_q8_0_f32(const char * cx, float * dst, const int ne,
+//                                   const int ne00, const int ne01, const int ne02, const int nb01, const int nb02, const int nb03) {
+//    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+//
+//    if (i >= ne) {
+//        return;
+//    }
+//
+//    const int64_t i03 = i/(ne00 * ne01 * ne02);
+//    const int64_t i02 = (i - i03*ne00*ne01*ne02) / (ne00*ne01);
+//    const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne00*ne01) / ne00;
+//    const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne00*ne01 - i01*ne00;
+//
+//    const block_q8_0 * q8 = (const block_q8_0 *)(cx + i01*nb01 + i02*nb02 + i03*nb03);
+//    const int ib = i00/QK8_0;
+//    const int iq = i00%QK8_0;
+//
+//    dst[i00*ne01 + i01 + i02*ne00*ne01 + i03*ne00*ne01*ne02] = __half2float(q8[ib].d)*q8[ib].qs[iq];
+//}
+
+static __global__ void k_transpose_q8_0(const char * cx, char * cdst,
+                                   const int ne10, const int ne11, const int ne12,
+                                   const int nb01, const int nb02, const int nb03,
+                                   const int nb11, const int nb12, const int nb13) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int64_t i13 = i/(ne10 * ne11 * ne12);
+    const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+
+    //const int64_t ne00 = ne11;
+    //const int64_t ne01 = ne10;
+    //const int64_t ne02 = ne12;
+    const int64_t i03 = i13;
+    const int64_t i02 = i12;
+    const int64_t i01 = i10; //(i - i03*ne00*ne01*ne02 - i02*ne00*ne01) / ne00;
+    const int64_t i00 = i11; //i - i03*ne00*ne01*ne02 - i02*ne00*ne01 - i01*ne00;
+
+    const block_q8_0 * q8 = (const block_q8_0 *)(cx + i01*nb01 + i02*nb02 + i03*nb03);
+    const int ib0 = i00/QK8_0;
+    const int iq0 = i00%QK8_0;
+
+    float xi = __half2float(q8[ib0].d)*q8[ib0].qs[iq0];
+    float amax = fabsf(xi);
+    amax = warp_reduce_max(amax);
+
+    //printf("%d, %d, %d: i = %ld, i11 = %ld i10 = %ld, xi = %g, amax = %g\n", blockDim.x, blockIdx.x, threadIdx.x, i, i11, i10, xi, amax);
+
+    float d = amax/127;
+    int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
+
+    block_q8_0 * dst = (block_q8_0 *)(cdst + i11*nb11 + i12*nb12 + i13*nb13);
+    dst[i10 / QK8_0].qs[i10 % QK8_0] = q;
+
+    if (threadIdx.x == 0) {
+        dst[i10 / QK8_0].d = __float2half(d);
+    }
+}
+
 static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
    const float * xi = (const float *) cxi;
    block_q8_0 * dsti = (block_q8_0 *) cdsti;
@@ -464,6 +525,32 @@ static void ggml_cpy_f16_f16_cuda(
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }

+static void transpose_q8_0(ggml_backend_cuda_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    auto stream = ctx.stream();
+    auto num_blocks = ggml_nelements(dst)/QK8_0;
+    k_transpose_q8_0<<<num_blocks, QK8_0, 0, stream>>>(
+            (const char *)src->data, (char *)dst->data,
+            dst->ne[0], dst->ne[1], dst->ne[2], src->nb[0], src->nb[2], src->nb[3],
+            dst->nb[1], dst->nb[2], dst->nb[3]);
+
+    //auto ne = ggml_nelements(dst);
+    //ggml_cuda_pool_alloc<float> dst_f32(ctx.pool(), ne);
+    //const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    //auto aux_src = *dst;
+    //aux_src.nb[0] = sizeof(float);
+    //aux_src.nb[1] = aux_src.nb[0]*aux_src.ne[0];
+    //aux_src.nb[2] = aux_src.nb[1]*aux_src.ne[1];
+    //aux_src.nb[3] = aux_src.nb[2]*aux_src.ne[2];
+    //cpy_q8_0_f32<<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+    //    ((const char *)src->data, dst_f32.get(), ne,
+    //     src->ne[1], src->ne[0], src->ne[2], src->nb[0], src->nb[2], src->nb[3]);
+    //CUDA_CHECK(cudaGetLastError());
+    //aux_src.type = GGML_TYPE_F32;
+    //ggml_cpy_f32_q8_0_cuda((const char *)dst_f32.get(), (char *)dst->data, ne, dst->ne[0], dst->ne[1], dst->ne[2],
+    //        aux_src.nb[0], aux_src.nb[1], aux_src.nb[2], aux_src.nb[3],
+    //        dst->ne[0], dst->ne[1], dst->ne[2], dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], stream);
+}
+
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
    const int64_t ne = ggml_nelements(src0);
    GGML_ASSERT(ne == ggml_nelements(src1));
@@ -522,9 +609,33 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (ggml_is_contiguous(src0) && ggml_are_same_shape(src0, src1)) {
+        if (src1->type == GGML_TYPE_F16) {
+            auto to_fp16 = ggml_get_to_fp16_cuda(src0->type);
+            if (to_fp16) {
+                to_fp16(src0->data, (half *)src1->data, ggml_nrows(src0), src0->ne[0], main_stream);
+            }
+        }
+        else if (src1->type == GGML_TYPE_F32) {
+            auto to_fp32 = ggml_get_to_fp32_cuda(src0->type);
+            if (to_fp32) {
+                to_fp32(src0->data, (float *)src1->data, ggml_nrows(src0), src0->ne[0], main_stream);
+            }
+        }
+        else if (src1->type == GGML_TYPE_BF16) {
+            auto to_bf16 = ggml_get_to_bf16_cuda(src0->type);
+            if (to_bf16) {
+                to_bf16(src0->data, (nv_bfloat16 *)src1->data, ggml_nrows(src0), src0->ne[0], main_stream);
+            }
+        }
+    } else if (ggml_are_same_shape(src0, src1) && src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        transpose_q8_0(ctx, src0, src1);
    } else {
        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
                ggml_type_name(src0->type), ggml_type_name(src1->type));
+        fprintf(stderr, "%s: %ld x %ld x %ld; %zu x %zu %zu -> %ld x %ld x %ld; %zu x %zu x %zu\n", __func__,
+                src0->ne[0], src0->ne[1], src0->ne[2], src0->nb[1], src0->nb[2], src0->nb[3],
+                src1->ne[0], src1->ne[1], src1->ne[2], src1->nb[1], src1->nb[2], src1->nb[3]);
        GGML_ABORT("fatal error");
    }
 }
@@ -559,9 +670,27 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
            return (void*) cpy_f32_f16<cpy_1_f16_f16>;
    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
            return (void*) cpy_f32_f16<cpy_1_f16_f32>;
-    } else {
-        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
-                ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ABORT("fatal error");
+    } else if (ggml_is_contiguous(src0) && ggml_are_same_shape(src0, src1)) {
+        if (src1->type == GGML_TYPE_F16) {
+            auto to_fp16 = ggml_get_to_fp16_cuda(src0->type);
+            if (to_fp16) return (void*)to_fp16;
+        }
+        else if (src1->type == GGML_TYPE_F32) {
+            auto to_fp32 = ggml_get_to_fp32_cuda(src0->type);
+            if (to_fp32) return (void*)to_fp32;
+        }
+        else if (src1->type == GGML_TYPE_BF16) {
+            auto to_bf16 = ggml_get_to_bf16_cuda(src0->type);
+            if (to_bf16) return (void*)to_bf16;
+        }
    }
+    else if (ggml_are_same_shape(src0, src1) && src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        return (void *)transpose_q8_0;
+    }
+    fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+            ggml_type_name(src0->type), ggml_type_name(src1->type));
+    fprintf(stderr, "%s: %ld x %ld x %ld; %zu x %zu %zu -> %ld x %ld x %ld; %zu x %zu x %zu\n", __func__,
+                src0->ne[0], src0->ne[1], src0->ne[2], src0->nb[1], src0->nb[2], src0->nb[3],
+                src1->ne[0], src1->ne[1], src1->ne[2], src1->nb[1], src1->nb[2], src1->nb[3]);
+    GGML_ABORT("fatal error");
 }