happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #2 from bluebread/sf/deepseek-ocr 

mtmd: DeepseekOCR Implement DeepSeek3B-MoE-A570M (LM component)
This commit is contained in:
Saba Fallah 2025-11-17 11:27:59 +01:00 committed by GitHub
parent 97e0907c5b 13dc6fb305
commit b32bb5e7da
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 13 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
convert_hf_to_gguf.py
View File

@ -7100,7 +7100,7 @@ class DeepseekV2Model(TextModel):
else: else:
# note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group) # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
self.hparams["num_key_value_heads"] = 1 self.hparams["num_key_value_heads"] = 1
super().set_gguf_parameters() super().set_gguf_parameters()
hparams = self.hparams hparams = self.hparams
kv_lora_rank = hparams["q_lora_rank"] if hparams["q_lora_rank"] is not None else 512 kv_lora_rank = hparams["q_lora_rank"] if hparams["q_lora_rank"] is not None else 512

18
examples/eval-callback/eval-callback.cpp
View File

@ -74,19 +74,19 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
} }
} }
for (int64_t i3 = 0; i3 < ne[3]; i3++) { for (int64_t i3 = 0; i3 < ne[3]; i3++) {
LOG(" [\n"); LOG(" [\n");
for (int64_t i2 = 0; i2 < ne[2]; i2++) { for (int64_t i2 = 0; i2 < ne[2]; i2++) {
if (i2 == n && ne[2] > 2*n) { if (i2 == n && ne[2] > 2*n) {
LOG(" ..., \n"); LOG(" ..., \n");
i2 = ne[2] - n; i2 = ne[2] - n;
} }
LOG(" [\n"); LOG(" [\n");
for (int64_t i1 = 0; i1 < ne[1]; i1++) { for (int64_t i1 = 0; i1 < ne[1]; i1++) {
if (i1 == n && ne[1] > 2*n) { if (i1 == n && ne[1] > 2*n) {
LOG(" ..., \n"); LOG(" ..., \n");
i1 = ne[1] - n; i1 = ne[1] - n;
} }
LOG(" ["); LOG(" [");
for (int64_t i0 = 0; i0 < ne[0]; i0++) { for (int64_t i0 = 0; i0 < ne[0]; i0++) {
if (i0 == n && ne[0] > 2*n) { if (i0 == n && ne[0] > 2*n) {
LOG("..., "); LOG("..., ");
@ -98,10 +98,10 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
} }
LOG("],\n"); LOG("],\n");
} }
LOG(" ],\n"); LOG(" ],\n");
} }
LOG(" ]\n"); LOG(" ]\n");
LOG(" sum = %f\n", sum); LOG(" sum = %f\n", sum);
} }
// TODO: make this abort configurable/optional? // TODO: make this abort configurable/optional?
@ -136,7 +136,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str()); snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
} }
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, LOG("%s: %16s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_desc(t), t->name, ggml_type_name(t->type), ggml_op_desc(t),
src0->name, ggml_ne_string(src0).c_str(), src0->name, ggml_ne_string(src0).c_str(),
src1 ? src1_str : "", src1 ? src1_str : "",

5
src/models/deepseek2.cpp
View File

@ -47,6 +47,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// self_attention // self_attention
if (is_ocr) { if (is_ocr) {
const int n_embed_head = hparams.n_embd / hparams.n_head(); const int n_embed_head = hparams.n_embd / hparams.n_head();
const int ocr_rope_type = GGML_ROPE_TYPE_NEOX;
GGML_ASSERT(n_embed_head == n_embd_head_k && n_embed_head == n_embd_head_v); GGML_ASSERT(n_embed_head == n_embd_head_k && n_embed_head == n_embd_head_v);
ggml_tensor * Qcur = NULL; ggml_tensor * Qcur = NULL;
@ -65,8 +66,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embed_head, n_head, n_tokens); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embed_head, n_head, n_tokens);
GGML_ASSERT(fabs(freq_base - 10000.0) < 1e-4); GGML_ASSERT(fabs(freq_base - 10000.0) < 1e-4);
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_embed_head, rope_type, 0, freq_base, 1, 0, 1, 0, 0); Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);
Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_embed_head, rope_type, 0, freq_base, 1, 0, 1, 0, 0); Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);
cb(Qcur, "q_pe", il); cb(Qcur, "q_pe", il);
cb(Kcur, "k_pe", il); cb(Kcur, "k_pe", il);