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 spec decoding warmup #906

Open
wants to merge 9 commits into
base: habana_main
Choose a base branch
from
Open

Fix spec decoding warmup #906

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
23cb4f6
spec warmup
yangw1234 Mar 11, 2025
60e94cb
fix lint
yangw1234 Mar 11, 2025
7cebbdc
fix mypy
yangw1234 Mar 11, 2025
1b15825
fix mypy
yangw1234 Mar 11, 2025
892a9d6
revert back example
yangw1234 Mar 11, 2025
8d9dbd0
skip proposer warmup
yangw1234 Mar 12, 2025
23f42f8
add comments
yangw1234 Mar 12, 2025
a7f19fa
revert back examples
yangw1234 Mar 13, 2025
519ef61
remove one line
yangw1234 Mar 13, 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
5 changes: 5 additions & 0 deletions vllm/spec_decode/hpu_draft_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,11 @@ def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)

self.indices_of_seq_with_bonus_tokens = None
# TODO: Currently, we skip warmup for TP1DraftModelRunner
# because in spec_decode_worker determine_num_available_blocks()
# is not called, so that warmup will fail. Simply adding this call
# does not work since other proposers do not implement this method.
self.model_runner.skip_warmup = True

@torch.inference_mode()
def execute_model(
Expand Down
31 changes: 28 additions & 3 deletions vllm/worker/hpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1892,10 +1892,33 @@ def warmup_scenario(self,
profiler.start()
for _ in range(times):
inputs = self.prepare_model_input(seqs)

additional_inputs = {}
if self.model_config.hf_config.model_type in ("medusa",
"mlp_speculator",
"eagle",
"deepseek_mtp"):
input_tokens = inputs.input_tokens
assert input_tokens is not None
bs = input_tokens.shape[0]
seq_len = input_tokens.shape[1]
hidden_size = self.model_config.get_hidden_size()

previous_hidden_states = torch.zeros(
(bs, seq_len, hidden_size),
device=input_tokens.device,
dtype=self.model_config.dtype)
additional_inputs = {
"previous_hidden_states": previous_hidden_states
}

is_single_step = \
self.vllm_config.scheduler_config.num_scheduler_steps == 1
if is_prompt or is_single_step:
self.execute_model(inputs, kv_caches, warmup_mode=True)
self.execute_model(inputs,
kv_caches,
warmup_mode=True,
**additional_inputs)
else: # decode with multi-step
inputs = dataclasses.replace(inputs,
is_first_multi_step=True,
Expand All @@ -1904,15 +1927,17 @@ def warmup_scenario(self,
kv_caches,
warmup_mode=True,
num_steps=2,
seqs=seqs)
seqs=seqs,
**additional_inputs)
inputs = dataclasses.replace(inputs,
is_first_multi_step=False,
is_last_step=True)
self.execute_model(inputs,
kv_caches,
warmup_mode=True,
num_steps=2,
seqs=seqs)
seqs=seqs,
**additional_inputs)
torch.hpu.synchronize()
if profiler:
profiler.step()
Expand Down
11 changes: 11 additions & 0 deletions vllm/worker/model_runner_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,3 +258,14 @@ def __init__(

def __getattr__(self, attr):
return getattr(self.model_runner, attr)

def __setattr__(self, name, value):
Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@michalkuligowski The purpose of this class is to imitate the behavior of a regular model runner and delegate all methods, attributes access to the self.model_runner.

Adding __setattr__ here is to solve the problem when setting a attr in base and use it later. E.g.

model_wrapper = ModelRunnerWrapperBase(hpu_model_runner)
model_wrapper.mem_margin = some_value
model_wrapper.mem_margin

model_wrapper.mem_margin = some_value this line will set a attr to the base class if we do not have __setattr__, but when we want to get the value by model_wrapper.mem_margin, the base class will delegate the access to self.model_runner which does not have this property and causing an error.

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can this be moved to hpu model runner?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No, move it to hpu model runner won't work, because the __setatt__ happens at wrapper class.

"""
Ensure that setting the 'model_runner' attribute
does not delegate to model_runner
"""

if name == "model_runner":
object.__setattr__(self, name, value)
else:
setattr(self.model_runner, name, value)