ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-16 13:58:09 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

CUDA: backwards pass for misc. ops, add tests (llama/11257)

Browse Source 
* CUDA: backwards pass for misc. ops, add tests

* remove restrict from pointers
This commit is contained in:
Johannes Gäßler
2025-01-16 16:43:38 +01:00
committed by Georgi Gerganov
parent db6383094c
commit de49024e49
17 changed files with 683 additions and 306 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
ggml/src/ggml.c
View File

@ -3454,12 +3454,14 @@ struct ggml_tensor * ggml_soft_max_ext(
return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
}
// ggml_soft_max_back
// ggml_soft_max_ext_back
static struct ggml_tensor * ggml_soft_max_back_impl(
static struct ggml_tensor * ggml_soft_max_ext_back_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
float scale,
float max_bias,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
@ -3467,21 +3469,28 @@ static struct ggml_tensor * ggml_soft_max_back_impl(
result->src[0] = a;
result->src[1] = b;
memcpy((float *) result->op_params + 0, &scale, sizeof(float));
memcpy((float *) result->op_params + 1, &max_bias, sizeof(float));
return result;
}
struct ggml_tensor * ggml_soft_max_back(
struct ggml_tensor * ggml_soft_max_ext_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b) {
return ggml_soft_max_back_impl(ctx, a, b, false);
struct ggml_tensor * b,
float scale,
float max_bias) {
return ggml_soft_max_ext_back_impl(ctx, a, b, scale, max_bias, false);
}
struct ggml_tensor * ggml_soft_max_back_inplace(
struct ggml_tensor * ggml_soft_max_ext_back_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b) {
return ggml_soft_max_back_impl(ctx, a, b, true);
struct ggml_tensor * b,
float scale,
float max_bias) {
return ggml_soft_max_ext_back_impl(ctx, a, b, scale, max_bias, true);
}
// ggml_rope
@ -5080,10 +5089,10 @@ struct ggml_tensor * ggml_cross_entropy_loss_back(
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c) {
GGML_ASSERT(ggml_are_same_shape(a, b));
GGML_ASSERT(ggml_is_scalar(c));
GGML_ASSERT(ggml_is_scalar(a));
GGML_ASSERT(ggml_are_same_shape(b, c));
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
struct ggml_tensor * result = ggml_dup_tensor(ctx, b);
result->op = GGML_OP_CROSS_ENTROPY_LOSS_BACK;
result->src[0] = a;
@ -5262,7 +5271,7 @@ static void ggml_sub_or_set(
}
static void ggml_compute_backward(
struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, bool * grads_needed) {
struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, const bool * grads_needed) {
struct ggml_tensor * tensor = cgraph->nodes[i];
struct ggml_tensor * grad = ggml_graph_get_grad(cgraph, tensor);
@ -5406,7 +5415,7 @@ static void ggml_compute_backward(
if (src0_needs_grads) {
float eps;
memcpy(&eps, tensor->op_params, sizeof(float));
ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, src0, grad, eps));
ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, grad, src0, eps));
}
} break;
case GGML_OP_MUL_MAT: {
@ -5589,7 +5598,13 @@ static void ggml_compute_backward(
} break;
case GGML_OP_SOFT_MAX: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_back(ctx, grad, tensor));
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (const float *) tensor->op_params + 0, sizeof(float));
memcpy(&max_bias, (const float *) tensor->op_params + 1, sizeof(float));
ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_ext_back(ctx, grad, tensor, scale, max_bias));
}
GGML_ASSERT((!src1 || !src1_needs_grads) && "backward pass for softmax mask not implemented");
} break;
@ -5630,7 +5645,7 @@ static void ggml_compute_backward(
const int32_t d1 = ggml_get_op_params_i32(tensor, 5);
const bool is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, src0, grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, grad, src0, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
}
} break;
case GGML_OP_POOL_2D: {
@ -5673,7 +5688,7 @@ static void ggml_compute_backward(
} break;
case GGML_UNARY_OP_SILU: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, src0, grad));
ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, grad, src0));
}
} break;
case GGML_UNARY_OP_EXP: {
@ -5690,7 +5705,7 @@ static void ggml_compute_backward(
} break;
case GGML_OP_CROSS_ENTROPY_LOSS: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, src0, src1, grad));
ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, grad, src0, src1));
}
GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
} break;