ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-16 05:48:09 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

llama: Add support for RWKV v7 architecture (llama/12412)

Browse Source 
* ggml: Add op l2_norm

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* ggml: Add op rwkv_wkv7

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: Add support for RWKV7 and ARWKV7 models

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: fix inference with RWKV6Qwen2

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: add more (a)rwkv7 variants in size

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* Apply code-format changes

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* fix MUSA build

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

* llama: fix shape error with rwkv using llama-parallel

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>

---------

Signed-off-by: Molly Sophia <mollysophia379@gmail.com>
This commit is contained in:
Molly Sophia
2025-03-18 07:27:50 +08:00
committed by Georgi Gerganov
parent cfc2560e41
commit 52c4c03b0a
21 changed files with 1751 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
ggml/src/ggml.c
View File

@ -929,6 +929,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"RMS_NORM",
"RMS_NORM_BACK",
"GROUP_NORM",
"L2_NORM",
"MUL_MAT",
"MUL_MAT_ID",
@ -977,6 +978,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"ADD_REL_POS",
"RWKV_WKV6",
"GATED_LINEAR_ATTN",
"RWKV_WKV7",
"UNARY",
@ -996,7 +998,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"OPT_STEP_ADAMW",
};
static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@ -1026,6 +1028,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"rms_norm(x)",
"rms_norm_back(x)",
"group_norm(x)",
"l2_norm(x)",
"X*Y",
"X[i]*Y",
@ -1074,6 +1077,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"add_rel_pos(x)",
"rwkv_wkv6(k, v, r, tf, td, s)",
"gated_linear_attn(k, v, q, gate, s)",
"rwkv_wkv7(r, w, k, v, a, b, s)",
"unary(x)",
@ -1093,7 +1097,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"adamw(x)",
};
static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@ -2686,6 +2690,37 @@ struct ggml_tensor * ggml_group_norm_inplace(
return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
}
// ggml_l2_norm
static struct ggml_tensor * ggml_l2_norm_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params_f32(result, 0, eps);
result->op = GGML_OP_L2_NORM;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_l2_norm(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps) {
return ggml_l2_norm_impl(ctx, a, eps, false);
}
struct ggml_tensor * ggml_l2_norm_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float eps) {
return ggml_l2_norm_impl(ctx, a, eps, true);
}
// ggml_mul_mat
static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
@ -4720,6 +4755,54 @@ struct ggml_tensor * ggml_gated_linear_attn(
return result;
}
// ggml_rwkv_wkv7
struct ggml_tensor * ggml_rwkv_wkv7(
struct ggml_context * ctx,
struct ggml_tensor * r,
struct ggml_tensor * w,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * state) {
GGML_ASSERT(ggml_is_contiguous(r));
GGML_ASSERT(ggml_is_contiguous(w));
GGML_ASSERT(ggml_is_contiguous(k));
GGML_ASSERT(ggml_is_contiguous(v));
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(ggml_is_contiguous(b));
GGML_ASSERT(ggml_is_contiguous(state));
const int64_t S = k->ne[0];
const int64_t H = k->ne[1];
const int64_t n_tokens = k->ne[2];
const int64_t n_seqs = state->ne[1];
{
GGML_ASSERT(w->ne[0] == S && w->ne[1] == H && w->ne[2] == n_tokens);
GGML_ASSERT(k->ne[0] == S && k->ne[1] == H && k->ne[2] == n_tokens);
GGML_ASSERT(v->ne[0] == S && v->ne[1] == H && v->ne[2] == n_tokens);
GGML_ASSERT(a->ne[0] == S && a->ne[1] == H && a->ne[2] == n_tokens);
GGML_ASSERT(b->ne[0] == S && b->ne[1] == H && b->ne[2] == n_tokens);
GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs);
}
// concat output and new_state
const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
result->op = GGML_OP_RWKV_WKV7;
result->src[0] = r;
result->src[1] = w;
result->src[2] = k;
result->src[3] = v;
result->src[4] = a;
result->src[5] = b;
result->src[6] = state;
return result;
}
// ggml_unary
static struct ggml_tensor * ggml_unary_impl(