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 all commits
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
26 changes: 23 additions & 3 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -523,15 +523,15 @@ def set_gguf_parameters(self):
self.gguf_writer.add_context_length(n_ctx)
logger.info(f"gguf: context length = {n_ctx}")

if (n_embd := self.find_hparam(["hidden_size", "n_embd"], optional=True)) is not None:
if (n_embd := self.find_hparam(["hidden_size", "n_embd", "dim"], optional=True)) is not None:
self.gguf_writer.add_embedding_length(n_embd)
logger.info(f"gguf: embedding length = {n_embd}")

if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
if (n_ff := self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
logger.info(f"gguf: feed forward length = {n_ff}")

if (n_head := self.find_hparam(["num_attention_heads", "n_head"], optional=True)) is not None:
if (n_head := self.find_hparam(["num_attention_heads", "n_head", "n_heads"], optional=True)) is not None:
self.gguf_writer.add_head_count(n_head)
logger.info(f"gguf: head count = {n_head}")

Expand Down Expand Up @@ -3907,6 +3907,26 @@ def _xlmroberta_set_vocab(self) -> None:
self.gguf_writer.add_add_eos_token(True)


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

def set_gguf_parameters(self):
self.gguf_writer.add_layer_norm_eps(1e-12)
logger.info("gguf: layer norm epsilon = 1e-12")
super().set_gguf_parameters()

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("distilbert."):
name = name[11:]

# These layers act as MLM head, so we don't need them
if name.startswith("vocab_"):
return []

return super().modify_tensors(data_torch, name, bid)


@ModelBase.register("RobertaModel", "RobertaForSequenceClassification")
class RobertaModel(BertModel):
model_arch = gguf.MODEL_ARCH.BERT
Expand Down
9 changes: 9 additions & 0 deletions gguf-py/gguf/tensor_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,6 +169,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.q_proj_no_perm", # llama-custom
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.layer.{bid}.attention.q_lin", # distillbert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq", # internlm2
Expand All @@ -183,6 +184,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.k_proj_no_perm", # llama-custom
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.layer.{bid}.attention.k_lin", # distillbert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"transformer.h.{bid}.attn.k", # refact
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
Expand All @@ -197,6 +199,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe olmo2 phimoe
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.layer.{bid}.attention.v_lin", # distillbert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"transformer.h.{bid}.attn.v", # refact
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
Expand All @@ -217,6 +220,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.linear_attn", # deci
"layers.{bid}.attention.wo", # llama-pth
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.layer.{bid}.attention.out_lin", # distillbert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"model.layers.{bid}.self_attn.dense", # persimmon
Expand All @@ -237,6 +241,7 @@ class TensorNameMap:
# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"transformer.layer.{bid}.sa_layer_norm", # distillbert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
"transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx
Expand Down Expand Up @@ -313,6 +318,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.up_proj", # llama-hf refact nemotron olmo2
"layers.{bid}.feed_forward.w3", # llama-pth
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.layer.{bid}.ffn.lin1", # distillbert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"transformer.h.{bid}.mlp.linear_3", # refact
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
Expand Down Expand Up @@ -396,6 +402,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.down_proj", # llama-hf nemotron olmo2
"layers.{bid}.feed_forward.w2", # llama-pth
"encoder.layer.{bid}.output.dense", # bert
"transformer.layer.{bid}.ffn.lin2", # distillbert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"model.layers.{bid}.mlp.dense_4h_to_h", # persimmon
Expand Down Expand Up @@ -457,6 +464,7 @@ class TensorNameMap:

MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"transformer.layer.{bid}.output_layer_norm", # distillbert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
"encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2
Expand Down Expand Up @@ -827,6 +835,7 @@ class TensorNameMap:
MODEL_TENSOR.CLS: (
"classifier", # jina
"classifier.dense", # roberta
"pre_classifier", # distillbert
),

MODEL_TENSOR.CLS_OUT: (
Expand Down
8 changes: 5 additions & 3 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2114,7 +2114,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
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}, 0);
type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);

if (arch == LLM_ARCH_BERT) {
pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, n_ctx_train}, 0);
Expand Down Expand Up @@ -5885,8 +5885,10 @@ struct llm_build_bert : public llm_graph_context {
inpL = build_inp_embd(model.tok_embd);

// token types are hardcoded to zero ("Sentence A")
ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
inpL = ggml_add(ctx0, inpL, type_row0);
if (model.type_embd) {
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) {
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
}
Expand Down
Loading