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.

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

[fix] update context estimate interface #1194

Merged
merged 1 commit into from
Oct 18, 2023
Merged

[fix] update context estimate interface #1194

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
57 changes: 24 additions & 33 deletions engines/python/setup/djl_python/transformers_neuronx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -99,55 +99,44 @@ def load_hf_model(self):
revision=self.revision,
low_cpu_mem_usage=True)

def get_model_specific_kwargs(self, model_type):
model_kwargs = {
"batch_size": self.batch_size,
"amp": self.amp,
'tp_degree': self.tensor_parallel_degree,
"n_positions": self.n_positions,
"unroll": self.unroll
}
if model_type == "llama":
model_kwargs[
'context_length_estimate'] = self.context_length_estimate
return model_kwargs

def load_inf2_model_from_disk(self, model_type, load_path):
if not self.load_split_model:
logging.info(f"Saving INF2 model to {load_path} ...")
save_pretrained_split(self.model, load_path)
model_kwargs = self.get_model_specific_kwargs(model_type)
if self.load_in_8bit:
neuron_config = NeuronConfig()
neuron_config.quant = QuantizationConfig(quant_dtype='s8',
dequant_dtype=self.amp)
return MODEL_TYPE_TO_MODEL[model_type].from_pretrained(
load_path,
batch_size=self.batch_size,
amp=self.amp,
tp_degree=self.tensor_parallel_degree,
n_positions=self.n_positions,
neuron_config=neuron_config,
context_length_estimate=self.context_length_estimate,
unroll=self.unroll)
load_path, neuron_config=neuron_config, **model_kwargs)
return MODEL_TYPE_TO_MODEL[model_type].from_pretrained(
load_path,
batch_size=self.batch_size,
amp=self.amp,
tp_degree=self.tensor_parallel_degree,
context_length_estimate=self.context_length_estimate,
n_positions=self.n_positions,
unroll=self.unroll)
load_path, **model_kwargs)

def load_inf2_model_from_memory(self, model_type):
model_kwargs = self.get_model_specific_kwargs(model_type)
if self.load_in_8bit:
neuron_config = NeuronConfig()
neuron_config.quant = QuantizationConfig(quant_dtype='s8',
dequant_dtype=self.amp)
model = MODEL_TYPE_TO_MODEL[model_type](
self.model.config,
batch_size=self.batch_size,
amp=self.amp,
tp_degree=self.tensor_parallel_degree,
context_length_estimate=self.context_length_estimate,
n_positions=self.n_positions,
neuron_config=neuron_config,
unroll=self.unroll)
self.model.config, neuron_config=neuron_config, **model_kwargs)
else:
model = MODEL_TYPE_TO_MODEL[model_type](
self.model.config,
batch_size=self.batch_size,
amp=self.amp,
tp_degree=self.tensor_parallel_degree,
context_length_estimate=self.context_length_estimate,
n_positions=self.n_positions,
unroll=self.unroll)
model = MODEL_TYPE_TO_MODEL[model_type](self.model.config,
**model_kwargs)
model.load_state_dict_low_memory(self.model.state_dict())
return model

Expand Down Expand Up @@ -176,7 +165,8 @@ def load_model(self, model_type):
self.model.to_neuron()
os.chdir(path)
elapsed = time.time() - start
logging.info(f"SysHealth: LLM sharding and compilation latency: {elapsed} secs")
logging.info(
f"SysHealth: LLM sharding and compilation latency: {elapsed} secs")

def initialize(self, properties):
# Neuron recommendation for transformersneuronx speedup
Expand Down Expand Up @@ -220,7 +210,8 @@ def initialize(self, properties):
self.low_cpu_mem_usage = True
if "context_length_estimate" in properties:
# expect input like [256, 1024, 2048]
self.context_length_estimate = json.loads(properties.get("context_length_estimate"))
self.context_length_estimate = json.loads(
properties.get("context_length_estimate"))
model_config = AutoConfig.from_pretrained(self.model_id_or_path,
revision=self.revision)
if model_config.model_type not in SUPPORTED_MODEL_TYPES:
Expand Down
Loading