Skip to content

Commit 075d1cb

Browse files
makaveli10makaveli10
committed
Fix out_prod; vulkan ci issues
1 parent b0c5b5b commit 075d1cb

File tree

5 files changed

+15
-45
lines changed

5 files changed

+15
-45
lines changed

‎examples/training/finetune.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -93,4 +93,4 @@ int main(int argc, char ** argv) {
9393
llama_backend_free();
9494

9595
return 0;
96-
}
96+
}

‎ggml/src/ggml-cuda/ggml-cuda.cu

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3202,7 +3202,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
32023202
}
32033203
} break;
32043204
case GGML_OP_OUT_PROD:
3205-
// return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
32063205
return op->type == GGML_TYPE_F32;
32073206
case GGML_OP_GET_ROWS:
32083207
{

‎ggml/src/ggml-cuda/out-prod.cu

Lines changed: 12 additions & 38 deletions
Original file line numberDiff line numberDiff line change
@@ -12,51 +12,46 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
1212
const bool src0_is_quantized = (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16);
1313
const bool src1_is_quantized = (src1->type != GGML_TYPE_F32 && src1->type != GGML_TYPE_F16);
1414

15-
// if (src0_is_quantized || src1_is_quantized) {
16-
// printf("DEBUG: OUT_PROD with quantized tensors - src0_quantized=%d, src1_quantized=%d\n",
17-
// src0_is_quantized, src1_is_quantized);
18-
// fflush(stdout);
19-
// }
20-
21-
// GGML_ASSERT(src0->type == GGML_TYPE_F32);
22-
// GGML_ASSERT(src1->type == GGML_TYPE_F32);
23-
2415
GGML_ASSERT(dst->type == GGML_TYPE_F32);
2516

17+
cudaStream_t stream = ctx.stream();
18+
ggml_cuda_pool & pool = ctx.pool();
19+
2620
// temp buffers
2721
float * src0_f32 = nullptr;
2822
float * src1_f32 = nullptr;
2923
bool allocated_src0 = false;
3024
bool allocated_src1 = false;
31-
cudaStream_t stream = ctx.stream();
25+
ggml_cuda_pool_alloc<float> src0_alloc(pool);
26+
ggml_cuda_pool_alloc<float> src1_alloc(pool);
3227

3328
if (src0_is_quantized) {
34-
const size_t src0_size = ggml_nelements(src0) * sizeof(float);
35-
CUDA_CHECK(cudaMallocAsync(&src0_f32, src0_size, stream));
29+
const size_t src0_size = ggml_nelements(src0);
30+
src0_alloc.alloc(src0_size);
31+
src0_f32 = src0_alloc.ptr;
3632
allocated_src0 = true;
3733

3834
// Dequantize
3935
auto dequantize_fn = ggml_get_to_fp32_cuda(src0->type);
4036
if (dequantize_fn) {
4137
dequantize_fn(src0->data, src0_f32, ggml_nelements(src0), stream);
4238
} else {
43-
CUDA_CHECK(cudaFreeAsync(src0_f32, stream));
4439
GGML_ABORT("Unsupported quant type for src0");
4540
}
4641
} else {
4742
src0_f32 = (float *) src0->data;
4843
}
4944

5045
if (src1_is_quantized) {
51-
const size_t src1_size = ggml_nelements(src1) * sizeof(float);
52-
CUDA_CHECK(cudaMallocAsync(&src1_f32, src1_size, stream));
46+
const size_t src1_size = ggml_nelements(src1);
47+
src1_alloc.alloc(src1_size);
48+
src1_f32 = src1_alloc.ptr;
5349
allocated_src1 = true;
5450

5551
auto dequantize_fn = ggml_get_to_fp32_cuda(src1->type);
5652
if (dequantize_fn) {
57-
dequantize_fn(src1->data, src1_f32, ggml_nelements(src0), stream);
53+
dequantize_fn(src1->data, src1_f32, ggml_nelements(src1), stream);
5854
} else {
59-
CUDA_CHECK(cudaFreeAsync(src1_f32, stream));
6055
GGML_ABORT("Unsupported quant type for src1");
6156
}
6257
} else {
@@ -74,9 +69,6 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
7469
GGML_ASSERT(ne2 == src1->ne[2]);
7570
GGML_ASSERT(ne3 == src1->ne[3]);
7671

77-
// const float * src0_d = (const float *) src0->data;
78-
// const float * src1_d = (const float *) src1->data;
79-
8072
// Use dequantized data
8173
const float * src0_d = src0_f32;
8274
const float * src1_d = src1_f32;
@@ -89,28 +81,21 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
8981

9082
CUBLAS_CHECK(cublasSetStream(handle, stream));
9183

92-
// const int64_t lda = nb01 / sizeof(float);
9384
const int64_t lda = allocated_src0 ? ne00 : (nb01 / sizeof(float));
9485
const int64_t ldc = nb1 / sizeof(float);
9586

9687
const bool src1_T = ggml_is_transposed(src1);
9788
const cublasOperation_t src1_cublas_op = src1_T ? CUBLAS_OP_N : CUBLAS_OP_T;
98-
// const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float);
9989
const int64_t ldb = allocated_src1 ?
10090
(src1_T ? ne10 : ne11) :
10191
((src1_T ? nb10 : nb11) / sizeof(float));
10292

103-
// GGML_ASSERT( (src1_T ? nb11 : nb10) == sizeof(float));
10493
// Only assert for non dequantized src1
10594
if (!allocated_src1) {
10695
GGML_ASSERT((src1_T ? nb11 : nb10) == sizeof(float));
10796
}
10897

10998
// data strides in dimensions 2/3
110-
// const size_t s02 = nb02 / sizeof(float);
111-
// const size_t s03 = nb03 / sizeof(float);
112-
// const size_t s12 = nb12 / sizeof(float);
113-
// const size_t s13 = nb13 / sizeof(float);
11499
const size_t s02 = allocated_src0 ? (ne00 * ne01) : nb02 / sizeof(float);
115100
const size_t s03 = allocated_src0 ? (ne00 * ne01 * ne02): nb03 / sizeof(float);
116101
const size_t s12 = allocated_src1 ? (ne10 * ne11) : nb12 / sizeof(float);
@@ -134,15 +119,4 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
134119
}
135120
}
136121

137-
if (allocated_src0) {
138-
CUDA_CHECK(cudaFreeAsync(src0_f32, stream));
139-
// printf("DEBUG: Freed dequantized src0 buffer\n");
140-
}
141-
if (allocated_src1) {
142-
CUDA_CHECK(cudaFreeAsync(src1_f32, stream));
143-
// // printf("DEBUG: Freed dequantized src1 buffer\n");
144-
}
145-
146-
// printf("DEBUG: CUDA OUT_PROD completed successfully\n");
147-
fflush(stdout);
148122
}

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp

Lines changed: 1 addition & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -9364,9 +9364,6 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
93649364
break;
93659365

93669366
case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
9367-
// std::cerr << "*** GGML_VK_BUILD_GRAPH: CROSS_ENTROPY_LOSS_BACK case hit, calling ggml_vk_cross_entropy_loss_back" << std::endl;
9368-
// std::cout << "*** GGML_VK_BUILD_GRAPH: CROSS_ENTROPY_LOSS_BACK case hit, calling ggml_vk_cross_entropy_loss_back" << std::endl;
9369-
// fflush(stdout); fflush(stderr);
93709367
ggml_vk_cross_entropy_loss_back(ctx, compute_ctx, src0, src1, src2, node, dryrun);
93719368

93729369
break;
@@ -10692,7 +10689,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
1069210689
case GGML_OP_SUB:
1069310690
case GGML_OP_MUL:
1069410691
case GGML_OP_DIV:
10695-
return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
10692+
return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
1069610693
(op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16) &&
1069710694
(op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
1069810695
case GGML_OP_SILU_BACK:

‎src/llama-lora-training.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -290,7 +290,7 @@ struct llama_adapter_lora * llama_lora_training_init(
290290
bool llama_opt_param_filter_lora(const struct ggml_tensor * tensor, void * userdata) {
291291
(void) userdata; // Unused param
292292

293-
if (!tensor || !tensor->name) {
293+
if (!tensor) {
294294
return false;
295295
}
296296

0 commit comments

Comments
 (0)