Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml: add ggml_rope_set_offset (+ metal support) (#27120)

Browse Source 
* add params

* cpu kernel

* metal kernel

* add test backend ops

* gate other backends

* ggml: (cuda) support ggml_rope_set_offset (#27121)

* rm cuda supports_op guard, fix webgpu clang-format

* ggml: support ggml_rope_set_offset on vulkan (#27344)

* ggml: support ggml_rope_set_offset on vulkan

* remove inplace optimization
This commit is contained in:
Xuan-Son Nguyen
2026-08-19 14:04:57 +02:00
committed by GitHub
parent 77acca437f
commit fe8156f789
19 changed files with 272 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml.c
+17 -1
View File
@@ -4200,7 +4200,7 @@ static struct ggml_tensor * ggml_rope_impl(
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
int32_t params[16] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
memcpy(params + 5, &freq_base, sizeof(float));
memcpy(params + 6, &freq_scale, sizeof(float));
memcpy(params + 7, &ext_factor, sizeof(float));
@@ -4212,6 +4212,8 @@ static struct ggml_tensor * ggml_rope_impl(
} else {
memset(params + 11, 0, sizeof(int32_t) * GGML_MROPE_SECTIONS);
}
params[15] = 0; // n_offs, set via ggml_rope_set_offset()
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE;
@@ -4422,6 +4424,20 @@ struct ggml_tensor * ggml_rope_multi_back(
result->op = GGML_OP_ROPE_BACK;
return result;
}
struct ggml_tensor * ggml_rope_set_offset(
struct ggml_tensor * a,
int n_offs) {
GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
GGML_ASSERT(n_offs >= 0);
const int32_t mode = ggml_get_op_params_i32(a, 2);
GGML_ASSERT(mode != GGML_ROPE_TYPE_VISION);
ggml_set_op_params_i32(a, 15, n_offs);
return a;
}
// ggml_clamp
struct ggml_tensor * ggml_clamp(