Skip to content
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.

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

Address Scalings in Dolomite Conversion #8

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft
Changes from all commits
Commits
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
34 changes: 28 additions & 6 deletions src/instructlab/dolomite/hf_models/model_conversion/llama.py
Original file line number Diff line number Diff line change
Expand Up @@ -229,6 +229,9 @@ def export_to_huggingface_llama(
config.num_key_value_heads,
config.n_embd // config.n_head,
AttentionHeadType(config.attention_head_type),
m_emb=config.m_emb,
m_residual=config.m_residual,
# m_width=config.m_width,
)

SafeTensorsWeightsManager.save_state_dict(state_dict, save_path)
Expand Down Expand Up @@ -285,11 +288,25 @@ def _export_state_dict_to_huggingface(
num_key_value_heads: int,
head_dim: int,
attention_head_type: AttentionHeadType,
m_residual: float = None,
m_emb: float = None,
m_width: float = None,
) -> None:
if m_residual is None:
m_residual = 1.
if m_emb is None:
m_emb = 1.

# NOTE: this will not work since the norms are tied
# has_m_width = False
# if m_width is None:
# has_m_width = True
# m_width = 1.

state_dict = {
"model.embed_tokens.weight": safetensors_weight_manager.get_tensor(
"transformer.wte.weight"
),
) * m_emb,
"model.norm.weight": safetensors_weight_manager.get_tensor(
"transformer.ln_f.weight"
),
Expand All @@ -298,7 +315,12 @@ def _export_state_dict_to_huggingface(
if safetensors_weight_manager.has_tensor("lm_head.weight"):
state_dict["lm_head.weight"] = safetensors_weight_manager.get_tensor(
"lm_head.weight"
)
) / m_width
# elif has_m_width:
# # int this we cannot tie
# state_dict["lm_head.weight"] = safetensors_weight_manager.get_tensor(
# "transformer.wte.weight"
# ) / m_width

for layer_idx in range(num_layers):
state_dict[f"model.layers.{layer_idx}.input_layernorm.weight"] = (
Expand Down Expand Up @@ -332,13 +354,13 @@ def _export_state_dict_to_huggingface(
state_dict[f"model.layers.{layer_idx}.mlp.down_proj.weight"] = (
safetensors_weight_manager.get_tensor(
f"transformer.h.{layer_idx}.mlp.c_proj.weight"
)
) * m_residual
)
if f"transformer.h.{layer_idx}.mlp.c_proj.bias" in safetensors_weight_manager:
state_dict[f"model.layers.{layer_idx}.mlp.down_proj.bias"] = (
safetensors_weight_manager.get_tensor(
f"transformer.h.{layer_idx}.mlp.c_proj.bias"
)
) * m_residual
)

query_weight, key_weight, value_weight = (
Expand Down Expand Up @@ -376,12 +398,12 @@ def _export_state_dict_to_huggingface(
safetensors_weight_manager.get_tensor(
f"transformer.h.{layer_idx}.attn.c_proj.weight"
)
)
) * m_residual
if f"transformer.h.{layer_idx}.attn.c_proj.bias" in safetensors_weight_manager:
state_dict[f"model.layers.{layer_idx}.self_attn.o_proj.bias"] = (
safetensors_weight_manager.get_tensor(
f"transformer.h.{layer_idx}.attn.c_proj.bias"
)
)
) * m_residual

return state_dict
Loading