vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Optimize RWKV6 Operator Naming and Implement Multi-core CPU/ SYCL Acceleration (#10133)

Browse source 
* rwkv6: rename to wkv6

* rwkv6: support avx2 avx512 armv8 armv9

* rwkv6: update cuda file name

* rwkv6: rename params

* wkv on sycl

* sycl: add some ops

* sycl: Enhance OP support judgment

* wkv6: drop armv9 and tranfer to GGML style

ggml-ci

* sync : ggml

* update the function to use appropriate types

* fix define error

* Update ggml/src/ggml-cpu.c

* add appropriate asserts

* move element-wise functions outside

* put the declaration outside the loop

* rewrite to be more inline with the common pattern for distributing threads

* use recommended way GGML_TENSOR_LOCALS

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Diego Devesa <slarengh@gmail.com>
Co-authored-by: Plamen Minev <pacominev@gmail.com>
Co-authored-by: Yuri Khrustalev <ykhrustalev@users.noreply.github.com>
Co-authored-by: Meng, Hengyu <airdldl@163.com>
This commit is contained in:
Zhiyuan Li 2024-11-07 18:19:10 +11:00 • committed by GitHub
parent 5c333e0140
commit 3bcd40b3c5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
22 changed files with 1977 additions and 1027 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/llama.cpp
View file

@ -7011,7 +7011,7 @@ static const std::map<llm_tensor, llm_tensor_info> llm_tensor_info_mapping = {
{LLM_TENSOR_TIME_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
{LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV}},
{LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
{LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_ATTN_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@ -7127,7 +7127,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
ggml_tensor * C = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs);
op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C);
} break;
case GGML_OP_RWKV_WKV:
case GGML_OP_RWKV_WKV6:
{
// FIXME
const int64_t S = 123;
@ -7140,7 +7140,7 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
ggml_tensor * tf = w;
ggml_tensor * td = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 1, S, H, n_tokens);
ggml_tensor * state = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, S, n_seqs, S, H);
op_tensor = ggml_rwkv_wkv(ctx, k, v, r, tf, td, state);
op_tensor = ggml_rwkv_wkv6(ctx, k, v, r, tf, td, state);
} break;
default:
GGML_ABORT("%s: missing test for op %s for tensor %s", __func__, ggml_op_name(op), w->name);
@ -10083,7 +10083,7 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
v = ggml_transpose(ctx, v);
r = ggml_transpose(ctx, r);
struct ggml_tensor * wkv_output = ggml_rwkv_wkv(ctx, k, v, r, layer->time_mix_first, w, *wkv_state);
struct ggml_tensor * wkv_output = ggml_rwkv_wkv6(ctx, k, v, r, layer->time_mix_first, w, *wkv_state);
cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0);
*wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));