8000 Add support for DistilBert by huydt84 · Pull Request #13907 · ggml-org/llama.cpp · GitHub
[go: up one dir, main page]
Skip to content

Add support for DistilBert #13907

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
May 30, 2025
Merged

Add support for DistilBert #13907

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
14d21fa
add distilbert
huydt-bti May 30, 2025
6d80075
Merge branch 'master' into huydt/distill-bert
huydt-bti May 30, 2025
2b2cf6d
small fixes
huydt-bti May 30, 2025
50e068d
add note for LLM_ARCH_DISTIL_BERT
huydt-bti May 30, 2025
913100b
Use MODEL_ARCH.BERT for DistilBert
huydt-bti May 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Use MODEL_ARCH.BERT for DistilBert
  • Loading branch information
@huydt-bti
huydt-bti committed May 30, 2025
commit 913100bd57f8811a43a12b7ea8e24df7b8ad5739
2 changes: 1 addition & 1 deletion convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3909,7 +3909,7 @@ def _xlmroberta_set_vocab(self) -> None:

@ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
class DistilBertModel(BertModel):
model_arch = gguf.MODEL_ARCH.DISTIL_BERT
model_arch = gguf.MODEL_ARCH.BERT

def set_gguf_parameters(self):
self.gguf_writer.add_layer_norm_eps(1e-12)
Expand Down
18 changes: 0 additions & 18 deletions gguf-py/gguf/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,7 +289,6 @@ class MODEL_ARCH(IntEnum):
STARCODER = auto()
REFACT = auto()
BERT = auto()
DISTIL_BERT = auto()
NOMIC_BERT = auto()
NOMIC_BERT_MOE = auto()
JINA_BERT_V2 = auto()
Expand Down Expand Up @@ -570,7 +569,6 @@ class MODEL_TENSOR(IntEnum):
MODEL_ARCH.STARCODER: "starcoder",
MODEL_ARCH.REFACT: "refact",
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.DISTIL_BERT: "distil-bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
MODEL_ARCH.NOMIC_BERT_MOE: "nomic-bert-moe",
MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2",
Expand Down Expand Up @@ -1049,22 +1047,6 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.CLS,
MODEL_TENSOR.CLS_OUT,
],
MODEL_ARCH.DISTIL_BERT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_EMBD_NORM,
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_OUT_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.LAYER_OUT_NORM,
MODEL_TENSOR.CLS,
MODEL_TENSOR.CLS_OUT,
],
MODEL_ARCH.NOMIC_BERT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_EMBD_NORM,
Expand Down
20 changes: 0 additions & 20 deletions src/llama-arch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_STARCODER, "starcoder" },
{ LLM_ARCH_REFACT, "refact" },
{ LLM_ARCH_BERT, "bert" },
{ LLM_ARCH_DISTIL_BERT, "distil-bert" },
{ LLM_ARCH_NOMIC_BERT, "nomic-bert" },
{ LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
{ LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
Expand Down Expand Up @@ -463,25 +462,6 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_CLS_OUT, "cls.output" },
},
},
{
LLM_ARCH_DISTIL_BERT,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_TOKEN_TYPES, "token_types" }, // not used
{ LLM_TENSOR_POS_EMBD, "position_embd" },
{ LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_CLS, "cls" },
{ LLM_TENSOR_CLS_OUT, "cls.output" },
},
},
{
LLM_ARCH_NOMIC_BERT,
{
Expand Down
1 change: 0 additions & 1 deletion src/llama-arch.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,6 @@ enum llm_arch {
LLM_ARCH_STARCODER,
LLM_ARCH_REFACT,
LLM_ARCH_BERT,
LLM_ARCH_DISTIL_BERT,
LLM_ARCH_NOMIC_BERT,
LLM_ARCH_NOMIC_BERT_MOE,
LLM_ARCH_JINA_BERT_V2,
Expand Down
13 changes: 4 additions & 9 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -679,7 +679,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.f_max_alibi_bias = 8.0f;
} break;
case LLM_ARCH_BERT:
case LLM_ARCH_DISTIL_BERT:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn);
Expand Down Expand Up @@ -2111,14 +2110,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_BERT:
case LLM_ARCH_DISTIL_BERT:
case LLM_ARCH_NOMIC_BERT:
case LLM_ARCH_NOMIC_BERT_MOE:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);

if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_DISTIL_BERT) {
if (arch == LLM_ARCH_BERT) {
pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}, 0);

cls = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
Expand Down Expand Up @@ -2162,7 +2160,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);

if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_DISTIL_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, 0);
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
Expand Down Expand Up @@ -5891,7 +5889,7 @@ struct llm_build_bert : public llm_graph_context {
ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
inpL = ggml_add(ctx0, inpL, type_row0);
}
if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_DISTIL_BERT) {
if (model.arch == LLM_ARCH_BERT) {
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
}
cb(inpL, "inp_embd", -1);
Expand Down Expand Up @@ -6008,7 +6006,7 @@ struct llm_build_bert : public llm_graph_context {
0.0f,
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
cb(cur, "ffn_moe_out", il);
} else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_DISTIL_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
} else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
cur = build_ffn(cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
NULL, NULL, NULL,
Expand Down Expand Up @@ -13191,7 +13189,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,

switch (arch) {
case LLM_ARCH_BERT:
case LLM_ARCH_DISTIL_BERT:
case LLM_ARCH_JINA_BERT_V2:
case LLM_ARCH_NOMIC_BERT:
case LLM_ARCH_NOMIC_BERT_MOE:
Expand Down Expand Up @@ -13298,7 +13295,6 @@ llm_graph_result_ptr llama_model::build_graph(
llm = std::make_unique<llm_build_refact>(*this, params, gf);
} break;
case LLM_ARCH_BERT:
case LLM_ARCH_DISTIL_BERT:
case LLM_ARCH_JINA_BERT_V2:
case LLM_ARCH_NOMIC_BERT:
case LLM_ARCH_NOMIC_BERT_MOE:
Expand Down Expand Up @@ -13666,7 +13662,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_GROK:
case LLM_ARCH_DBRX:
case LLM_ARCH_BERT:
case LLM_ARCH_DISTIL_BERT:
case LLM_ARCH_NOMIC_BERT:
case LLM_ARCH_NOMIC_BERT_MOE:
case LLM_ARCH_STABLELM:
Expand Down
Loading
0