ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : fix YARN + add tests + add asserts (#7617)

Browse source 
* tests : add rope tests

ggml-ci

* ggml : fixes (hopefully)

ggml-ci

* tests : add non-cont tests

ggml-ci

* cuda : add asserts for rope/norm + fix DS2

ggml-ci

* ggml : assert contiguousness

* tests : reduce RoPE tests

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-05-29 20:17:31 +03:00 • committed by GitHub
parent cce3dcffc5
commit fb76ec31a9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 167 additions and 104 deletions
Download patch file Download diff file Expand all files Collapse all files

52
llama.cpp
View file

@ -11187,46 +11187,69 @@ struct llm_build_context {
}
// split into {n_head * n_embd_head_qk_nope, n_tokens}
struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_element_size(q) * hparams.n_embd_head_k, ggml_element_size(q) * hparams.n_embd_head_k * n_head, 0);
struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k),
ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
0);
cb(q_nope, "q_nope", il);
// and {n_head * n_embd_head_qk_rope, n_tokens}
struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_element_size(q) * hparams.n_embd_head_k, ggml_element_size(q) * hparams.n_embd_head_k * n_head, ggml_element_size(q) * n_embd_head_qk_nope);
struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k),
ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
ggml_row_size(q->type, n_embd_head_qk_nope));
cb(q_pe, "q_pe", il);
// {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
struct ggml_tensor * compressed_kv_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
cb(compressed_kv_pe, "compressed_kv_pe", il);
struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
cb(kv_pe_compresseed, "kv_pe_compresseed", il);
// split into {kv_lora_rank, n_tokens}
struct ggml_tensor * compressed_kv = ggml_view_2d(ctx0, compressed_kv_pe, kv_lora_rank, n_tokens, compressed_kv_pe->nb[1], 0);
cb(compressed_kv, "compressed_kv", il);
struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
kv_pe_compresseed->nb[1],
0);
cb(kv_compressed, "kv_compressed", il);
// and {n_embd_head_qk_rope, n_tokens}
struct ggml_tensor * k_pe = ggml_view_2d(ctx0, compressed_kv_pe, n_embd_head_qk_rope, n_tokens, compressed_kv_pe->nb[1], ggml_element_size(compressed_kv_pe)*kv_lora_rank);
struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
kv_pe_compresseed->nb[1],
kv_pe_compresseed->nb[1],
ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
cb(k_pe, "k_pe", il);
compressed_kv = llm_build_norm(ctx0, compressed_kv, hparams,
kv_compressed = ggml_cont(ctx0, kv_compressed); // TODO: the CUDA backend does not support non-contiguous norm
kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
model.layers[il].attn_kv_a_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(compressed_kv, "compressed_kv", il);
cb(kv_compressed, "kv_compressed", il);
// {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, compressed_kv);
struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
cb(kv, "kv", il);
// split into {n_head * n_embd_head_qk_nope, n_tokens}
struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, ggml_element_size(kv) * (n_embd_head_qk_nope + hparams.n_embd_head_v), ggml_element_size(kv) * n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v), 0);
struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
0);
cb(k_nope, "k_nope", il);
// and {n_head * n_embd_head_v, n_tokens}
struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, ggml_element_size(kv) * (n_embd_head_qk_nope + hparams.n_embd_head_v), ggml_element_size(kv) * n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v), ggml_element_size(kv) * n_embd_head_qk_nope);
struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
ggml_row_size(kv->type, (n_embd_head_qk_nope)));
cb(v_states, "v_states", il);
v_states = ggml_cont(ctx0, v_states);
cb(v_states, "v_states", il);
v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, ggml_element_size(kv) * hparams.n_embd_head_v * n_head, 0);
v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
0);
cb(v_states, "v_states", il);
q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
q_pe = ggml_rope_ext(
ctx0, q_pe, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
@ -11235,8 +11258,9 @@ struct llm_build_context {
cb(q_pe, "q_pe", il);
// shared RoPE key
k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
k_pe = ggml_rope_ext(
ctx0, ggml_view_3d(ctx0, k_pe, n_embd_head_qk_rope, 1, n_tokens, k_pe->nb[0], k_pe->nb[1], 0), inp_pos, nullptr,
ctx0, k_pe, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor_scaled, beta_fast, beta_slow
);