tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 19:55:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: Add support for RWKV v7 architecture (#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 GitHub
parent 60c902926c
commit 7dfad387e3
35 changed files with 2948 additions and 438 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
tests/test-backend-ops.cpp
View File

@ -1916,6 +1916,40 @@ struct test_gla : public test_case {
}
};
// GGML_OP_RWKV_WKV7
struct test_rwkv_wkv7 : public test_case {
const ggml_type type;
const int64_t head_count;
const int64_t head_size;
const int64_t n_seq_tokens;
const int64_t n_seqs;
std::string vars() override {
return VARS_TO_STR5(type, head_count, head_size, n_seq_tokens, n_seqs);
}
test_rwkv_wkv7(ggml_type type = GGML_TYPE_F32,
int64_t head_count = 32, int64_t head_size = 64, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
: type(type), head_count(head_count), head_size(head_size), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
const int64_t n_tokens = n_seq_tokens * n_seqs;
ggml_tensor * r = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
ggml_tensor * w = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
ggml_tensor * k = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
ggml_tensor * v = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
ggml_tensor * a = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
ggml_tensor * b = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
// Outputs may become NaN with long seqlen without these normalization
a = ggml_l2_norm(ctx, a, 1e-7F);
b = ggml_l2_norm(ctx, b, 1e-7F);
ggml_tensor * s = ggml_new_tensor(ctx, type, 2, std::vector<int64_t>{ head_size * head_size * head_count, n_seqs }.data());
ggml_tensor * out = ggml_rwkv_wkv7(ctx, r, w, k, v, a, b, s);
return out;
}
};
// GGML_OP_MUL_MAT
struct test_mul_mat : public test_case {
const ggml_type type_a;
@ -2972,6 +3006,32 @@ struct test_group_norm : public test_case {
}
};
// GGML_OP_L2_NORM
struct test_l2_norm : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const float eps;
std::string vars() override {
return VARS_TO_STR2(type, ne);
}
test_l2_norm(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {64, 64, 320, 1},
float eps = 1e-12f)
: type(type), ne(ne), eps(eps) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_name(a, "a");
ggml_tensor * out = ggml_l2_norm(ctx, a, eps);
ggml_set_name(out, "out");
return out;
}
};
// GGML_OP_ACC
struct test_acc : public test_case {
const ggml_type type;
@ -4006,8 +4066,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 5, 4, 3}, v, eps));
}
test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
test_cases.emplace_back(new test_l2_norm (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
}
test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
@ -4019,6 +4082,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 4));
test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 128, 4));
test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 1));
test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 1));
test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 4));
test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 128, 4));
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 1, 1));
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 1));
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));