Skip to content

inference optimizations #17

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

Open
TheJDen wants to merge 1 commit into
base: pipe
Choose a base branch
from
Open

inference optimizations #17

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
53 changes: 27 additions & 26 deletions deep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,27 +23,28 @@ def clear_gpu_memory():
def create_and_initialize_model(hidden_size, intermediate_size, num_hidden_layers, ds_config):
from deepspeed import zero

with zero.Init(config_dict_or_path=ds_config):
num_attention_heads = max(1, hidden_size // 64) # Ensure divisibility
llama_config = LlamaConfig(
vocab_size=32000,
hidden_size=hidden_size,
intermediate_size=intermediate_size,
num_hidden_layers=num_hidden_layers,
num_attention_heads=num_attention_heads,
num_key_value_heads=num_attention_heads,
max_position_embeddings=2048,
rms_norm_eps=1e-6,
use_cache=True,
pad_token_id=0,
bos_token_id=1,
eos_token_id=2,
tie_word_embeddings=False,
activation_function="swiGLU"
)

model = LlamaForCausalLM(llama_config)
return model
# dont need ZeRO 3 for inference

num_attention_heads = max(1, hidden_size // 64) # Ensure divisibility
llama_config = LlamaConfig(
vocab_size=32000,
hidden_size=hidden_size,
intermediate_size=intermediate_size,
num_hidden_layers=num_hidden_layers,
num_attention_heads=num_attention_heads,
num_key_value_heads=num_attention_heads,
max_position_embeddings=2048,
rms_norm_eps=1e-6,
use_cache=True,
pad_token_id=0,
bos_token_id=1,
eos_token_id=2,
tie_word_embeddings=False,
activation_function="swiGLU",
)

model = LlamaForCausalLM(llama_config)
return torch.compile(model, mode="max-autotune") # opt 2

def main():
# Remove manual initialization of the distributed process group
Expand Down Expand Up @@ -90,11 +91,11 @@ def main():
print(f"Total parameter count: {total_params:,}")

try:
model_engine, _, _, _ = deepspeed.initialize(
model=model,
model_parameters=model.parameters(),
config=ds_config
)
model_engine = deepspeed.init_inference(model, # opt 1
tensor_parallel={"tp_size": num_gpus},
dtype=torch.bfloat16,
replace_with_kernel_inject=False, # deepspeed doesn't support swiGLU yet
)

# Move tokenizer inputs to the correct device
input_text = "DeepSpeed is"
Expand Down