Skip to content

LLM inference in C/C++

C++ ggml

ggml : fix ggml_clamp (#27644)

Georgi Gerganov6036c635e2d1
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 32462d79a7..5f6774a630 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -1724,6 +1724,19 @@ extern "C" {
struct ggml_tensor * a,
int n_past);
+ GGML_API struct ggml_tensor * ggml_clamp(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ float min,
+ float max);
+
+ // in-place, returns view(a)
+ GGML_API struct ggml_tensor * ggml_clamp_inplace(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ float min,
+ float max);
+
GGML_API struct ggml_tensor * ggml_soft_max(
struct ggml_context * ctx,
struct ggml_tensor * a);
@@ -1990,14 +2003,6 @@ extern "C" {
struct ggml_tensor * a,
int n_offs);
- // clamp
- // in-place, returns view(a)
- GGML_API struct ggml_tensor * ggml_clamp(
- struct ggml_context * ctx,
- struct ggml_tensor * a,
- float min,
- float max);
-
// im2col
// converts data into a format that effectively results in a convolution when combined with matrix multiplication
GGML_API struct ggml_tensor * ggml_im2col(
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index 3bda9abbe0..a71838eafc 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -40,6 +40,7 @@ bool ggml_op_can_inplace(enum ggml_op op) {
case GGML_OP_SILU_BACK:
case GGML_OP_RMS_NORM:
case GGML_OP_RMS_NORM_BACK:
+ case GGML_OP_CLAMP:
case GGML_OP_SOFT_MAX:
case GGML_OP_SOFT_MAX_BACK:
return true;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 1a60fec791..e0b615c07e 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -4042,6 +4042,41 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace(
return ggml_diag_mask_zero_impl(ctx, a, n_past, true);
}
+// ggml_clamp
+
+static struct ggml_tensor * ggml_clamp_impl(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ float min,
+ float max,
+ bool inplace) {
+ struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+ float params[] = { min, max };
+ ggml_set_op_params(result, params, sizeof(params));
+
+ result->op = GGML_OP_CLAMP;
+ result->src[0] = a;
+
+ return result;
+}
+
+struct ggml_tensor * ggml_clamp(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ float min,
+ float max) {
+ return ggml_clamp_impl(ctx, a, min, max, false);
+}
+
+struct ggml_tensor * ggml_clamp_inplace(
+ struct ggml_context * ctx,
+ struct ggml_tensor * a,
+ float min,
+ float max) {
+ return ggml_clamp_impl(ctx, a, min, max, true);
+}
+
// ggml_soft_max
static struct ggml_tensor * ggml_soft_max_impl(
@@ -4438,25 +4473,6 @@ struct ggml_tensor * ggml_rope_set_offset(
return a;
}
-// ggml_clamp
-
-struct ggml_tensor * ggml_clamp(
- struct ggml_context * ctx,
- struct ggml_tensor * a,
- float min,
- float max) {
- // TODO: when implement backward, fix this:
- struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
- float params[] = { min, max };
- ggml_set_op_params(result, params, sizeof(params));
-
- result->op = GGML_OP_CLAMP;
- result->src[0] = a;
-
- return result;
-}
-
static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
}