apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-24 01:29:57 +08:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

[MTP] refactor MTP pre_process (#6358)

Browse Source
This commit is contained in:
周周周
2026-02-09 10:47:15 +08:00
committed by GitHub
parent 18e79dd660
commit 2b4748de4f
24 changed files with 411 additions and 533 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/layers/sample/ops/apply_penalty_multi_scores.py
+10 -10
View File
@@ -175,8 +175,8 @@ def apply_speculative_penalty_multi_scores(
min_dec_lens: paddle.Tensor,
eos_token_ids: paddle.Tensor,
seq_lens_this_time: paddle.Tensor,
output_padding_offset: paddle.Tensor,
output_cum_offsets: paddle.Tensor,
batch_id_per_token_output: paddle.Tensor,
cu_seqlens_q_output: paddle.Tensor,
max_len: int,
):
"""
@@ -200,8 +200,8 @@ def apply_speculative_penalty_multi_scores(
min_dec_lens,
eos_token_ids,
seq_lens_this_time,
output_padding_offset,
output_cum_offsets,
batch_id_per_token_output,
cu_seqlens_q_output,
max_len,
)
elif current_platform.is_xpu():
@@ -221,8 +221,8 @@ def apply_speculative_penalty_multi_scores(
min_dec_lens,
eos_token_ids,
seq_lens_this_time,
output_padding_offset,
output_cum_offsets,
batch_id_per_token_output,
cu_seqlens_q_output,
max_len,
)
@@ -242,8 +242,8 @@ def reasoning_phase_token_constraint(
step_idx: paddle.Tensor,
reasoning_allowed_tokens: paddle.Tensor,
reasoning_status: paddle.Tensor,
output_padding_offset: paddle.Tensor,
output_cum_offsets: paddle.Tensor,
batch_id_per_token_output: paddle.Tensor,
cu_seqlens_q_output: paddle.Tensor,
enable_thinking: paddle.Tensor,
think_end_id: int,
line_break_id: int,
@@ -263,8 +263,8 @@ def reasoning_phase_token_constraint(
step_idx,
reasoning_allowed_tokens,
reasoning_status,
output_padding_offset,
output_cum_offsets,
batch_id_per_token_output,
cu_seqlens_q_output,
enable_thinking,
think_end_id,
line_break_id,
fastdeploy/model_executor/layers/sample/sampler.py
+8 -8
View File
@@ -758,8 +758,8 @@ class SpeculativeSampler(nn.Layer):
sampling_metadata.min_dec_lens,
sampling_metadata.eos_token_ids,
share_inputs["seq_lens_this_time"],
share_inputs["output_padding_offset"],
share_inputs["output_cum_offsets"],
share_inputs["batch_id_per_token_output"],
share_inputs["cu_seqlens_q_output"],
max_model_len,
)
@@ -773,8 +773,8 @@ class SpeculativeSampler(nn.Layer):
share_inputs["step_idx"],
share_inputs["reasoning_allowed_tokens"],
share_inputs["reasoning_status"],
share_inputs["output_padding_offset"],
share_inputs["output_cum_offsets"],
share_inputs["batch_id_per_token_output"],
share_inputs["cu_seqlens_q_output"],
share_inputs["enable_thinking"],
self.think_end_id,
self.line_break_id,
@@ -794,7 +794,7 @@ class SpeculativeSampler(nn.Layer):
verify_scores, verify_tokens, actual_candidate_len = top_p_candidates(
probs,
sampling_metadata.top_p,
share_inputs["output_padding_offset"],
share_inputs["batch_id_per_token_output"],
self.speculative_max_candidate_len,
max_model_len,
)
@@ -816,7 +816,7 @@ class SpeculativeSampler(nn.Layer):
share_inputs["max_dec_len"],
sampling_metadata.eos_token_ids,
share_inputs["is_block_step"],
share_inputs["output_cum_offsets"],
share_inputs["cu_seqlens_q_output"],
actual_candidate_len,
share_inputs["actual_draft_token_num"],
sampling_metadata.top_p,
@@ -1135,8 +1135,8 @@ class MTPSampler(nn.Layer):
sampling_metadata.min_dec_lens,
sampling_metadata.eos_token_ids,
share_inputs["seq_lens_this_time"],
share_inputs["output_padding_offset"],
share_inputs["output_cum_offsets"],
share_inputs["batch_id_per_token_output"],
share_inputs["cu_seqlens_q_output"],
max_model_len,
)
probs = F.softmax(logits)