Skip to content

fix: read and set head_size and head_dim when convert Eagle checkpoint #3633

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

Closed
wants to merge 1 commit into from
Closed

fix: read and set head_size and head_dim when convert Eagle checkpoint #3633

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
14 changes: 13 additions & 1 deletion examples/eagle/convert_checkpoint.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -295,6 +295,8 @@ def copy(tensors):
args.n_positions = hf_config.max_position_embeddings
args.dtype = str(
hf_config.torch_dtype)[6:] if args.dtype == 'auto' else args.dtype
args.head_dim = hf_config.head_dim
args.head_size = args.head_dim

if args.eagle_model_dir is None:
hf_config_eagle = hf_config.eagle
Expand All @@ -305,6 +307,12 @@ def copy(tensors):
args.n_kv_head_eagle = hf_config_eagle['num_key_value_heads']
args.rms_norm_eps_eagle = hf_config_eagle['rms_norm_eps']
args.n_positions_eagle = hf_config_eagle['max_position_embeddings']
if 'head_dim' in hf_config_eagle:
args.head_dim_eagle = hf_config_eagle['head_dim']
else:
args.head_dim_eagle = args.n_embd_eagle // args.n_head_eagle

args.head_size_eagle = args.head_dim_eagle
else:
hf_config_eagle = LlamaConfig.from_pretrained(args.eagle_model_dir)
args.n_head_eagle = hf_config_eagle.num_attention_heads
Expand Down Expand Up @@ -370,6 +378,8 @@ def copy(tensors):
},
'use_parallel_embedding': args.use_parallel_embedding,
'embedding_sharding_dim': args.embedding_sharding_dim,
'head_dim': args.head_dim_eagle,
'head_size': args.head_size_eagle
}

config = {
Expand Down Expand Up @@ -402,7 +412,9 @@ def copy(tensors):
'max_draft_len': args.max_draft_len,
'num_eagle_layers': args.num_eagle_layers,
'max_non_leaves_per_layer': args.max_non_leaves_per_layer,
'eagle_net_config': eagle_net_config
'eagle_net_config': eagle_net_config,
'head_dim': args.head_dim,
'head_size': args.head_size
}

assert args.max_draft_len <= 256, "args.max_draft_len > 256 is not supported"
Expand Down
4 changes: 4 additions & 0 deletions tensorrt_llm/models/eagle/config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,6 +88,8 @@ def from_hugging_face(
n_positions = hf_config.max_position_embeddings
hidden_act = hf_config.hidden_act
dtype = str(hf_config.torch_dtype)[6:] if dtype == 'auto' else dtype
head_dim = hf_config.head_dim
head_size = hf_config.head_size

if speculative_config_or_dir is None:
hf_config_eagle = hf_config.eagle
Expand Down Expand Up @@ -143,6 +145,8 @@ def from_hugging_face(
},
'use_parallel_embedding': kwargs['use_parallel_embedding'],
'embedding_sharding_dim': kwargs['embedding_sharding_dim'],
'head_dim': head_dim,
'head_size': head_size
}

config = {
Expand Down