Skip to content

[Feature] support calculate entropy while samling #5348

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
Deleter-D wants to merge 1 commit into
base: feature/experimental_feature_20250908
Choose a base branch
from
Open

[Feature] support calculate entropy while samling #5348

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
117 changes: 117 additions & 0 deletions fastdeploy/model_executor/layers/sample/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,7 @@
top_k_top_p_sampling,
)
from fastdeploy.platforms import current_platform
from fastdeploy.utils import data_processor_logger
from fastdeploy.worker.output import LogprobsTensors, SamplerOutput

if current_platform.is_cuda():
Expand Down Expand Up @@ -223,6 +224,8 @@ def __init__(self, fd_config: FDConfig = None):
self.early_stopper = early_stopper_cls()
self.early_stopper.initialize(fd_config.parallel_config.max_num_seqs, fd_config.early_stop_config)

self.entropy_list = [[] for _ in range(fd_config.parallel_config.max_num_seqs)]

def apply_logits_processor(
self,
ids: int,
Expand Down Expand Up @@ -317,6 +320,40 @@ def gather_logprobs(

return LogprobsTensors(indices, top_logprobs, token_ranks)

def calculate_logits_entropy(self, logits, share_inputs, temperature):
real_bsz = share_inputs["seq_lens_this_time"].shape[0]
real_seq_lens = paddle.where(
share_inputs["seq_lens_encoder"][:real_bsz].squeeze(1) != 0,
paddle.ones([1], dtype="int32"),
share_inputs["seq_lens_this_time"].squeeze(1),
)

def get_entropy(logits):
a0 = logits - paddle.max(logits, axis=-1, keepdim=True)
ea0 = paddle.exp(a0)
z0 = paddle.sum(ea0, axis=-1, keepdim=True)
p0 = ea0 / z0
return paddle.sum(p0 * (paddle.log(z0) - a0), axis=-1)

batch_indices = paddle.arange(real_bsz, dtype="int32")
batch_id_per_token = paddle.repeat_interleave(batch_indices, real_seq_lens)
# print(f"[Sampler][entropy] batch_id_per_token: {batch_id_per_token}")
for i in range(logits.shape[0]):
if temperature[batch_id_per_token[i]] > 0 and temperature[batch_id_per_token[i]] != 1.0:
logits[i] = logits[i].scale_(1 / temperature[batch_id_per_token[i]])

entropy_tensor = get_entropy(logits)
entropy = entropy_tensor.tolist()

for i in range(real_bsz):
for _ in range(real_seq_lens[i]):
self.entropy_list[i].append(entropy.pop(0))
if share_inputs["stop_flags"][i] and len(self.entropy_list[i]) != 0:
data_processor_logger.info(
f"req_id: {share_inputs['req_ids'][i]}, entropy: {sum(self.entropy_list[i])/len(self.entropy_list[i])}"
)
self.entropy_list[i] = []

def forward_cuda(
self,
logits: paddle.Tensor,
Expand Down Expand Up @@ -374,6 +411,19 @@ def forward_cuda(
logprobs_tensors=logprobs_tensors,
)

# print(f"[Sampler] req_ids: {share_inputs['req_ids']}")
# print(f"[Sampler] next_tokens: {next_tokens}")
# print(f"[Sampler] seq_lens_this_time: {share_inputs['seq_lens_this_time']}")
# print(f"[Sampler] seq_lens_encoder: {share_inputs['seq_lens_encoder']}")
# print(f"[Sampler] logits: {logits}")
# print(f"[Sampler] temperature: {sampling_metadata.temperature}")
# print(f"[Sampler] stop_flags: {share_inputs['stop_flags']}")
# print(f"[Sampler] entropy_list: {self.entropy_list}")

self.calculate_logits_entropy(
paddle.clone(logits), sampling_metadata.share_inputs, sampling_metadata.temperature
)

return sampler_output


Expand All @@ -393,6 +443,7 @@ def __init__(self, fd_config: FDConfig):
self.speculative_max_candidate_len = fd_config.speculative_config.max_candidate_len
self.speculative_benchmark_mode = fd_config.speculative_config.benchmark_mode
self.speculative_tokens_num = fd_config.speculative_config.num_speculative_tokens
self.entropy_list = [[] for _ in range(fd_config.parallel_config.max_num_seqs)]

def pre_process(self, skip_idx_list: List[int] = []):
"""pre process before running"""
Expand Down Expand Up @@ -499,6 +550,58 @@ def gather_logprobs(

return LogprobsTensors(indices, top_logprobs, token_ranks)

def calculate_logits_entropy(self, logits, share_inputs, temperature):
# get accepted logits
real_bsz = share_inputs["seq_lens_this_time"].shape[0]
total_accepted_num = paddle.sum(share_inputs["accept_num"])
real_seq_lens = paddle.where(
share_inputs["seq_lens_encoder"][:real_bsz].squeeze(1) != 0,
paddle.ones([1], dtype="int32"),
share_inputs["seq_lens_this_time"].squeeze(1),
)
# print(f"[entropy] real_seq_lens: {real_seq_lens}")
seq_start_idx = paddle.concat([paddle.zeros([1], dtype="int32"), paddle.cumsum(real_seq_lens, dtype="int32")])
# print(f"[entropy] seq_start_idx: {seq_start_idx}")
repeated_starts = paddle.repeat_interleave(seq_start_idx[:-1], share_inputs["accept_num"][:real_bsz])
# print(f"[entropy] repeated_starts: {repeated_starts}")
offsets = paddle.concat([paddle.arange(share_inputs["accept_num"][i].item()) for i in range(real_bsz)]).astype(
"int32"
)
# print(f"[entropy] offsets: {offsets}")
accepted_idx = repeated_starts + offsets
# print(f"[entropy] accepted_idx: {accepted_idx}")

accepted_logits = paddle.empty([total_accepted_num, logits.shape[1]], dtype=logits.dtype)
# print(f"[entropy] accepted_logits shape: {accepted_logits.shape}")
for i in range(total_accepted_num):
accepted_logits[i] = logits[accepted_idx[i]]

def get_entropy(logits):
a0 = logits - paddle.max(logits, axis=-1, keepdim=True)
ea0 = paddle.exp(a0)
z0 = paddle.sum(ea0, axis=-1, keepdim=True)
p0 = ea0 / z0
return paddle.sum(p0 * (paddle.log(z0) - a0), axis=-1)

batch_indices = paddle.arange(share_inputs["accept_num"].shape[0], dtype="int32")
batch_id_per_token = paddle.repeat_interleave(batch_indices, share_inputs["accept_num"])
# print(f"[SpeculativeSampler][entropy] batch_id_per_token: {batch_id_per_token}")
for i in range(accepted_logits.shape[0]):
if temperature[batch_id_per_token[i]] > 0 and temperature[batch_id_per_token[i]] != 1.0:
accepted_logits[i] = accepted_logits[i].scale_(1 / temperature[batch_id_per_token[i]])

entropy_tensor = get_entropy(accepted_logits)
entropy = entropy_tensor.tolist()

for i in range(real_bsz):
for _ in range(share_inputs["accept_num"][i]):
self.entropy_list[i].append(entropy.pop(0))
if share_inputs["stop_flags"][i] and len(self.entropy_list[i]) != 0:
data_processor_logger.info(
f"req_id: {share_inputs['req_ids'][i]}, entropy: {sum(self.entropy_list[i])/len(self.entropy_list[i])}"
)
self.entropy_list[i] = []

def forward_cuda(
self,
logits: paddle.Tensor,
Expand All @@ -509,6 +612,7 @@ def forward_cuda(
reject_all_drafts: bool = False,
think_end_id: int = -1,
line_break_id: int = -1,
is_dummy_run: bool = False,
) -> paddle.Tensor:
logits = apply_speculative_penalty_multi_scores(
sampling_metadata.pre_token_ids,
Expand Down Expand Up @@ -631,6 +735,19 @@ def forward_cuda(
cu_batch_token_offset=share_inputs["cu_batch_token_offset"],
)

if not is_dummy_run:
# print(f"[SpeculativeSampler] req_ids: {share_inputs['req_ids']}")
# print(f"[SpeculativeSampler] accept_tokens: {share_inputs['accept_tokens']}")
# print(f"[SpeculativeSampler] accept_num: {share_inputs['accept_num']}")
# print(f"[SpeculativeSampler] seq_lens_this_time: {share_inputs['seq_lens_this_time']}")
# print(f"[SpeculativeSampler] seq_lens_encoder: {share_inputs['seq_lens_encoder']}")
# print(f"[SpeculativeSampler] logits: {logits}")
# print(f"[SpeculativeSampler] temperature: {sampling_metadata.temperature}")
# print(f"[SpeculativeSampler] stop_flags: {share_inputs['stop_flags']}")
# print(f"[SpeculativeSampler] entropy_list: {self.entropy_list}")

self.calculate_logits_entropy(logits, share_inputs, sampling_metadata.temperature)

return sampler_output


Expand Down
Loading