Files
xc-llm-ascend/vllm_ascend/sample/rejection_sampler.py

1011 lines
37 KiB
Python
Raw Normal View History

Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
# SPDX-License-Identifier: Apache-2.0
from typing import Optional
import torch
from vllm.triton_utils import HAS_TRITON, tl, triton
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.rejection_sampler import (GREEDY_TEMPERATURE,
generate_uniform_probs)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
from vllm_ascend.sample.sampler import apply_top_k_top_p
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
PLACEHOLDER_TOKEN_ID = -1
# Maximum number of speculative draft tokens allowed per request in a single
# step. This value is chosen to be large enough to handle typical use cases.
MAX_SPEC_LEN = 32
vectorcore_num = None
device_properties = None
if HAS_TRITON:
from triton.runtime import driver # type: ignore
device_properties = driver.active.utils.get_device_properties(
torch.npu.current_device())
vectorcore_num = device_properties['num_vectorcore']
#get vector core number in order for later tiling
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def apply_sampling_constraints(
logits: torch.Tensor, # [num_tokens, vocab_size]
cu_num_draft_tokens: torch.Tensor, # [batch_size]
sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
"""Process logits based on sampling metadata.
This function applies temperature scaling to the logits,
as well as top-k and top-p. For greedy decoding, it returns
the original logits.
Args:
logits: Input logits tensor to be processed.
cu_num_draft_tokens: Cumulative number of draft tokens.
sampling_metadata: Metadata containing sampling parameters such as
temperature and whether greedy sampling is used.
Returns:
torch.Tensor: Processed logits if non-greedy sampling is used,
otherwise returns the original logits.
"""
assert logits.ndim == 2
assert cu_num_draft_tokens.ndim == 1
if sampling_metadata.all_greedy:
return logits
num_tokens = logits.shape[0]
temperature = expand_batch_to_tokens(
sampling_metadata.temperature,
cu_num_draft_tokens,
num_tokens,
replace_from=GREEDY_TEMPERATURE,
replace_to=1,
)
# NOTE(woosuk): Update `logits` in place to avoid allocating a new tensor.
logits.div_(temperature.unsqueeze(-1))
# Get expanded top_k and top_p tensors.
top_k = None
if sampling_metadata.top_k is not None:
top_k = expand_batch_to_tokens(
sampling_metadata.top_k,
cu_num_draft_tokens,
num_tokens,
)
top_p = None
if sampling_metadata.top_p is not None:
top_p = expand_batch_to_tokens(
sampling_metadata.top_p,
cu_num_draft_tokens,
num_tokens,
)
# NOTE(woosuk): `apply_top_k_top_p` uses sorting to calculate the mask,
# which is slow for large vocab sizes. This may cause performance issues.
return apply_top_k_top_p(logits, top_k, top_p)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def rejection_sample(
# [num_tokens]
draft_token_ids: torch.Tensor,
# [batch_size]
num_draft_tokens: list[int],
max_spec_len: int,
# [batch_size]
cu_num_draft_tokens: torch.Tensor,
# [num_tokens, vocab_size]
draft_probs: Optional[torch.Tensor],
# [num_tokens, vocab_size]
target_probs: torch.Tensor,
# [batch_size, 1]
bonus_token_ids: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> torch.Tensor:
assert draft_token_ids.ndim == 1
assert draft_probs is None or draft_probs.ndim == 2
assert cu_num_draft_tokens.ndim == 1
assert target_probs.ndim == 2
batch_size = len(num_draft_tokens)
num_tokens = draft_token_ids.shape[0]
vocab_size = target_probs.shape[-1]
device = target_probs.device
assert draft_token_ids.is_contiguous()
assert draft_probs is None or draft_probs.is_contiguous()
assert target_probs.is_contiguous()
assert bonus_token_ids.is_contiguous()
assert target_probs.shape == (num_tokens, vocab_size)
# Create output buffer.
output_token_ids = torch.empty(
(batch_size, max_spec_len + 1),
dtype=torch.int32, # Consistent with SamplerOutput.sampled_token_ids.
device=device,
)
output_token_ids.fill_(PLACEHOLDER_TOKEN_ID)
if sampling_metadata.all_greedy:
is_greedy = None
else:
is_greedy = sampling_metadata.temperature == GREEDY_TEMPERATURE
if not sampling_metadata.all_random:
# Rejection sampling for greedy sampling requests.
target_argmax = target_probs.argmax(dim=-1)
if HAS_TRITON:
vec_len = batch_size
n = cu_num_draft_tokens.numel()
BLOCK_SIZE = 2
grid = triton.cdiv(n, BLOCK_SIZE)
if n >= vectorcore_num:
grid = vectorcore_num # Empirically tuned value
BLOCK_SIZE = triton.next_power_of_2(triton.cdiv(n, grid))
if min(num_draft_tokens) == 1 and max(
num_draft_tokens) == 1 and sampling_metadata.all_greedy:
rejection_greedy_sample_spec_len_1_triton[(grid, )](
output_token_ids,
draft_token_ids,
target_argmax,
bonus_token_ids,
vec_len,
BLOCK_SIZE=BLOCK_SIZE,
)
else:
rejection_greedy_sample_triton[(grid, )](
output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
target_argmax,
bonus_token_ids,
is_greedy,
vec_len,
max_spec_len,
BLOCK_SIZE=BLOCK_SIZE,
)
else:
if min(num_draft_tokens) == 1 and max(
num_draft_tokens) == 1 and sampling_metadata.all_greedy:
rejection_greedy_sample_spec_len_1_pytorch(
output_token_ids,
draft_token_ids,
target_argmax,
bonus_token_ids,
)
else:
rejection_greedy_sample_pytorch(
output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
target_argmax,
bonus_token_ids,
num_draft_tokens,
max_spec_len,
is_greedy,
)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
if sampling_metadata.all_greedy:
return output_token_ids
# Generate uniform probabilities for rejection sampling.
# [num_tokens]
uniform_probs = generate_uniform_probs(
num_tokens,
num_draft_tokens,
sampling_metadata.generators,
device,
)
# Sample recovered tokens for each position.
# [num_tokens]
recovered_token_ids = sample_recovered_tokens(
max_spec_len,
num_draft_tokens,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
sampling_metadata,
device,
)
# Rejection sampling for random sampling requests.
if HAS_TRITON:
rejection_random_sample_kernel[(batch_size, )](
output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
bonus_token_ids,
recovered_token_ids,
uniform_probs.to(torch.float32),
is_greedy,
max_spec_len,
vocab_size,
NO_DRAFT_PROBS=draft_probs is None,
)
else:
rejection_random_sample_pytorch(
output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
bonus_token_ids,
recovered_token_ids,
uniform_probs,
is_greedy,
max_spec_len,
vocab_size,
IS_NGRAM=draft_probs is None,
# num_warps=1,
)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
return output_token_ids
def expand_batch_to_tokens(
x: torch.Tensor, # [batch_size]
cu_num_tokens: torch.Tensor, # [batch_size]
num_tokens: int,
replace_from: int = 0,
replace_to: int = 0,
) -> torch.Tensor:
"""Expand [batch_size] tensor to [num_tokens] tensor based on the number of
tokens per batch in cu_num_tokens.
For example, if x = [a, b, c] and cu_num_tokens = [2, 5, 6], then
num_tokens = 6, and expanded_x = [a, a, b, b, b, c].
Args:
x: [batch_size] tensor to expand.
cu_num_tokens: [batch_size] tensor containing the cumulative number of
tokens per batch. Each element represents the total number of
tokens up to and including that batch.
num_tokens: Total number of tokens.
replace_from: int = 0
Value to be replaced if it is found in x.
replace_to: int = 0
Value to replace with when replace_from is found.
Returns:
expanded_x: [num_tokens] tensor.
"""
batch_size = x.shape[0]
assert cu_num_tokens.shape[0] == batch_size
expanded_x = x.new_empty(num_tokens)
if HAS_TRITON:
vec_len = batch_size
n = cu_num_tokens.numel()
BLOCK_SIZE = 2
grid = triton.cdiv(n, BLOCK_SIZE)
if n >= vectorcore_num:
grid = vectorcore_num
BLOCK_SIZE = triton.next_power_of_2(triton.cdiv(n, grid))
expand_kernel[(grid, )](
expanded_x,
x,
cu_num_tokens,
replace_from,
replace_to,
vec_len,
MAX_NUM_TOKENS=MAX_SPEC_LEN, # To avoid recompilation.
BLOCK_SIZE=BLOCK_SIZE,
)
else:
expand_pytorch(
expanded_x,
x,
cu_num_tokens,
replace_from,
replace_to,
MAX_NUM_TOKENS=MAX_SPEC_LEN, # To avoid recompilation.
)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
return expanded_x
def sample_recovered_tokens(
max_spec_len: int,
num_draft_tokens: list[int],
cu_num_draft_tokens: torch.Tensor,
draft_token_ids: torch.Tensor,
draft_probs: Optional[torch.Tensor],
target_probs: torch.Tensor,
sampling_metadata: SamplingMetadata,
device: torch.device,
) -> torch.Tensor:
batch_size = len(num_draft_tokens)
vocab_size = target_probs.shape[-1]
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
q = torch.empty(
(batch_size, vocab_size),
dtype=torch.float32,
device=device,
)
q.exponential_()
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
num_draft_tensor = torch.tensor(num_draft_tokens,
pin_memory=True).to(device,
non_blocking=True)
has_draft_mask = num_draft_tensor > 0
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
for i, generator in sampling_metadata.generators.items():
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
temp_q = torch.empty_like(q[i])
temp_q.exponential_(generator=generator)
q[i] = torch.where(has_draft_mask[i], temp_q, q[i])
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
recovered_token_ids = torch.empty_like(draft_token_ids)
if HAS_TRITON:
sample_recovered_tokens_kernel[(batch_size, max_spec_len)](
recovered_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
q,
vocab_size,
triton.next_power_of_2(vocab_size),
NO_DRAFT_PROBS=draft_probs is None,
SUB_BLOCK=4 * 1024,
# TODO: enable multibuffer when accuracy problem is solved.
multibuffer=False,
)
else:
sample_recovered_tokens_pytorch(
recovered_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
q,
vocab_size,
IS_NGRAM=draft_probs is None,
)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
return recovered_token_ids
def rejection_greedy_sample_spec_len_1_pytorch(
output_token_ids, # [batch_size, 2]
draft_token_ids, # [num_tokens]
target_argmax, # [num_tokens]
bonus_token_ids, # [batch_size]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
):
batch_size = output_token_ids.size(0)
num_tokens = draft_token_ids.size(0)
assert batch_size == num_tokens
accept_req_mask = draft_token_ids == target_argmax
output_token_ids[:, 0] = target_argmax
bonus_token_ids = bonus_token_ids.squeeze(1)
output_token_ids[:, 1] = torch.where(accept_req_mask, bonus_token_ids,
output_token_ids[:, 1])
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def rejection_greedy_sample_pytorch(
output_token_ids, # [batch_size, max_spec_len + 1]
cu_num_draft_tokens, # [batch_size]
draft_token_ids, # [num_tokens]
target_argmax, # [num_tokens]
bonus_token_ids, # [batch_size]
draft_tokens_per_req, # [batch_size], list
max_spec_len,
is_greedy=None, # [batch_size] or None
):
batch_size = output_token_ids.size(0)
num_tokens = draft_token_ids.size(0)
device = output_token_ids.device
draft_tokens_per_req = torch.tensor(draft_tokens_per_req).to(
device, non_blocking=True)
if is_greedy is None:
is_greedy = torch.ones(batch_size, dtype=torch.bool, device=device)
start_indices = cu_num_draft_tokens - draft_tokens_per_req
req_ids = torch.arange(batch_size, device=device)
token_req_ids = torch.repeat_interleave(req_ids, draft_tokens_per_req)
token_positions = torch.arange(
num_tokens, device=device) - start_indices[token_req_ids]
# Find the first mismatch position of each request.
mismatch_global = (draft_token_ids != target_argmax)
if max_spec_len == 0:
first_mismatch_pos_per_req = torch.zeros(batch_size,
dtype=torch.long,
device=device)
else:
# [bs, max_spec_len]
pos_matrix = torch.full((batch_size, max_spec_len),
-1,
dtype=torch.long,
device=device)
pos_matrix[token_req_ids, token_positions] = token_positions
mismatch_matrix = torch.full((batch_size, max_spec_len),
False,
dtype=torch.bool,
device=device)
mismatch_matrix[token_req_ids, token_positions] = mismatch_global
mismatch_positions = torch.where(mismatch_matrix, pos_matrix,
max_spec_len * 2)
first_mismatch_pos_per_req, _ = torch.min(mismatch_positions, dim=1)
no_mismatch_mask = (first_mismatch_pos_per_req == max_spec_len * 2)
first_mismatch_pos_per_req[no_mismatch_mask] = draft_tokens_per_req[
no_mismatch_mask]
# Copy matched target tokens into output.
copy_len = torch.minimum(first_mismatch_pos_per_req + 1,
draft_tokens_per_req)
copy_indices = torch.arange(max_spec_len + 1,
device=device).expand(batch_size, -1)
copy_mask = copy_indices < copy_len.unsqueeze(1)
greedy_mask = is_greedy.unsqueeze(1)
final_copy_mask = copy_mask & greedy_mask
global_idx = start_indices.unsqueeze(1) + copy_indices
output_token_ids[final_copy_mask] = target_argmax[
global_idx[final_copy_mask]].to(output_token_ids.dtype)
# Fill bonus token.
needs_bonus = is_greedy & (first_mismatch_pos_per_req
>= draft_tokens_per_req)
if torch.any(needs_bonus):
bonus_rows = torch.where(needs_bonus)[0]
bonus_cols = draft_tokens_per_req[bonus_rows]
bonus_token_ids = bonus_token_ids.squeeze(1)
output_token_ids[bonus_rows, bonus_cols] = bonus_token_ids[bonus_rows]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def rejection_random_sample_pytorch(
output_token_ids, # [batch_size, max_spec_len + 1]
cu_num_draft_tokens, # [batch_size]
draft_token_ids, # [num_tokens]
draft_probs, # [num_tokens, vocab_size] or None
target_probs, # [num_tokens, vocab_size]
bonus_token_ids, # [batch_size]
recovered_token_ids, # [num_tokens]
uniform_probs, # [num_tokens]
is_greedy, # [batch_size]
max_spec_len,
vocab_size,
IS_NGRAM=False,
):
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
"""
This function implements the Speculative Decoding rejection sampling step.
Instead of looping through each request and each token (which causes high
overhead), it uses a fully vectorized approach:
1. **Index Mapping**: Converts the flattened 1D token arrays into a 2D
[batch_size, max_draft_len] grid using 'cu_num_draft_tokens' to handle
variable-length sequences in the batch.
2. **Parallel Validation**: Calculates the acceptance condition
(target_prob / draft_prob >= uniform_sample) for ALL draft tokens
simultaneously across the entire batch.
3. **Short-circuit Simulation**: In the loop version, once a token is rejected,
subsequent tokens are ignored. Here, we simulate this by finding the
'first_reject_pos' using argmax on the rejection mask and creating a
'should_skip' mask for all indices after the first failure.
4. **Token Selection**: Uses 'torch.where' to select:
- Draft tokens (if accepted)
- Recovered tokens (at the point of first rejection)
- Bonus tokens (if all tokens in a sequence were accepted)
5. **Masking**: Ensures operations only apply to non-greedy requests and
within valid sequence lengths.
"""
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
batch_size = output_token_ids.shape[0]
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
device = output_token_ids.device
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
zero_cpu = torch.tensor([0], pin_memory=True)
zero_device = zero_cpu.to(device, non_blocking=True)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
cu_start = torch.cat([zero_device, cu_num_draft_tokens[:-1]])
cu_end = cu_num_draft_tokens
num_draft_per_batch = cu_end - cu_start
max_draft_len = max_spec_len
pos_indices_cpu = torch.arange(max_draft_len, pin_memory=True)
pos_indices = pos_indices_cpu.to(device, non_blocking=True)[None, :]
valid_mask = pos_indices < num_draft_per_batch[:, None]
global_token_indices = cu_start[:, None] + pos_indices
global_token_indices = global_token_indices.clamp(
0, draft_token_ids.shape[0] - 1)
draft_tokens = draft_token_ids[
global_token_indices] # [batch_size, max_draft_len]
if IS_NGRAM:
ones_cpu = torch.ones(1, pin_memory=True, dtype=torch.float32)
draft_token_probs = ones_cpu.to(
device, non_blocking=True).expand_as(draft_tokens)
else:
flat_indices = global_token_indices.flatten()
flat_draft_tokens = draft_tokens.flatten()
flat_draft_probs = draft_probs[flat_indices, flat_draft_tokens]
draft_token_probs = flat_draft_probs.view(batch_size, max_draft_len)
flat_indices = global_token_indices.flatten()
flat_draft_tokens = draft_tokens.flatten()
flat_target_probs = target_probs[flat_indices, flat_draft_tokens]
target_token_probs = flat_target_probs.view(batch_size, max_draft_len)
uniform_token_probs = uniform_probs[global_token_indices]
recovered_tokens = recovered_token_ids[global_token_indices]
zero_threshold_cpu = torch.tensor([0.0],
pin_memory=True,
dtype=torch.float32)
zero_threshold = zero_threshold_cpu.to(device, non_blocking=True)
acceptance_condition = (draft_token_probs > zero_threshold) & (
target_token_probs / draft_token_probs >= uniform_token_probs)
first_rejection = (~acceptance_condition) & valid_mask
default_pos_cpu = torch.full([batch_size, 1],
max_draft_len,
pin_memory=True)
default_pos = default_pos_cpu.to(device, non_blocking=True)
first_reject_pos = torch.where(
first_rejection.any(dim=1, keepdim=True),
first_rejection.float().argmax(dim=1, keepdim=True), default_pos)
pos_mask = pos_indices >= first_reject_pos
should_skip = pos_mask & valid_mask
final_acceptance = acceptance_condition & (~should_skip)
non_greedy_mask = ~is_greedy
update_mask = non_greedy_mask[:, None] & valid_mask & (~should_skip)
first_reject_mask = (pos_indices == first_reject_pos
) & valid_mask & non_greedy_mask[:, None]
final_update_mask = update_mask | first_reject_mask
final_tokens = torch.where(
first_reject_mask, recovered_tokens,
torch.where(final_acceptance, draft_tokens,
output_token_ids[:, :max_draft_len]))
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
output_token_ids[:, :max_draft_len] = torch.where(
final_update_mask, final_tokens, output_token_ids[:, :max_draft_len])
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
no_rejection = first_reject_pos.squeeze(1) >= num_draft_per_batch
should_add_bonus = non_greedy_mask & no_rejection
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
bonus_positions = num_draft_per_batch # [batch_size]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
seq_len = output_token_ids.shape[1]
all_positions_cpu = torch.arange(seq_len, pin_memory=True)
all_positions = all_positions_cpu.to(
device, non_blocking=True)[None, :] # [1, seq_len]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
batch_bonus_positions = bonus_positions[:, None] # [batch_size, 1]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
max_spec_len_cpu = torch.tensor([max_spec_len], pin_memory=True)
max_spec_len_device = max_spec_len_cpu.to(device, non_blocking=True)
valid_bonus_pos = bonus_positions < (max_spec_len_device + 1)
final_bonus_mask = should_add_bonus & valid_bonus_pos
bonus_pos_match = (all_positions == batch_bonus_positions)
bonus_pos_mask = bonus_pos_match & final_bonus_mask[:, None]
bonus_values_expanded = bonus_token_ids.view(-1, 1).expand(-1, seq_len)
output_token_ids[:] = torch.where(bonus_pos_mask, bonus_values_expanded,
output_token_ids)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def expand_pytorch(
output_ptr, # [num_tokens]
input_ptr, # [batch_size]
cu_num_tokens_ptr, # [batch_size]
replace_from,
replace_to,
MAX_NUM_TOKENS,
):
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
"""
This function broadcasts batch-level values (input_ptr) to token-level
positions (output_ptr) based on cumulative token offsets. It acts like
a "scatter" or "repeat_interleave" operation but with custom logic:
1. **Range Broadcasting**: It creates a boolean matrix 'in_range' of size
[num_tokens, batch_size] that identifies which batch index each token
belongs to by checking if the token index falls between cu_start and cu_end.
2. **Conditional Replacement**: Before expansion, it replaces specific values
(e.g., padding or special markers) in the input to prepare the data.
3. **Matrix-based Mapping**: It uses 'torch.einsum' to perform a weighted
sum that effectively "picks" the correct batch value for every token position
simultaneously, avoiding a Python loop over the batch.
"""
device = cu_num_tokens_ptr.device
batch_size = input_ptr.shape[0]
num_tokens = output_ptr.shape[0]
if batch_size == 0 or num_tokens == 0:
return
cu_start = torch.cat([
torch.tensor([0], pin_memory=True).to(device, non_blocking=True),
cu_num_tokens_ptr[:-1]
])
cu_end = cu_num_tokens_ptr
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
token_indices = torch.arange(num_tokens,
device=device)[:, None] # [num_tokens, 1]
cu_start_exp = cu_start[None, :] # [1, batch_size]
cu_end_exp = cu_end[None, :] # [1, batch_size]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
in_range = (token_indices >= cu_start_exp) & (token_indices < cu_end_exp)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
replaced_input = torch.where(input_ptr == replace_from, replace_to,
input_ptr).float()
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
token_values = torch.einsum("tb,b->t", in_range.float(), replaced_input)
needs_update = in_range.any(dim=1)
output_ptr[:] = torch.where(needs_update, token_values, output_ptr)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
def sample_recovered_tokens_pytorch(
output_token_ids, # [num_tokens]
cu_num_draft_tokens, # [batch_size]
draft_token_ids, # [num_tokens]
draft_probs, # [num_tokens, vocab_size] or None
target_probs, # [num_tokens, vocab_size]
q, # [batch_size, vocab_size]
vocab_size,
IS_NGRAM=False,
):
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
"""
When a draft token is rejected, we must sample a "recovered" token from
a modified distribution. This function calculates that distribution across
the entire flattened batch.
1. **Token-to-Batch Mapping**: Using the cumulative draft token counts, it
determines which request in the batch each token belongs to. This is
necessary because 'q' (normalization factor) is stored per-request.
2. **Probability Adjustment**:
- If N-GRAM: It zeroes out the draft token's probability in the target.
- If Probabilistic: It calculates max(0, target_probs - draft_probs)
as per the standard speculative decoding algorithm.
3. **Normalization & Sampling**: It divides the adjusted probabilities
by the normalization distribution 'q'. To remain vectorized, it
broadcasts 'q' from [batch_size, vocab] to [num_tokens, vocab].
4. **Argmax Selection**: It selects the best recovery token for every
position in one pass using torch.argmax.
"""
device = output_token_ids.device
num_tokens = output_token_ids.shape[0]
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
if num_tokens == 0:
return
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
cu_start = torch.cat([
torch.tensor([0], pin_memory=True).to(device, non_blocking=True),
cu_num_draft_tokens[:-1],
])
cu_end = cu_num_draft_tokens
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
token_indices = torch.arange(num_tokens, device=device) # [num_tokens]
token_indices_expanded = token_indices[:, None] # [num_tokens, 1]
cu_start_expanded = cu_start[None, :] # [1, batch_size]
cu_end_expanded = cu_end[None, :] # [1, batch_size]
in_range_mask = (token_indices_expanded >= cu_start_expanded) & (
token_indices_expanded < cu_end_expanded)
token_to_batch = torch.argmax(in_range_mask.int(), dim=1)
has_match = in_range_mask.any(dim=1)
token_to_batch = torch.where(has_match, token_to_batch, 0)
if IS_NGRAM:
token_indices = torch.arange(num_tokens, device=device)
modified_target_probs = target_probs.clone()
modified_target_probs[token_indices, draft_token_ids] = 0
prob = modified_target_probs
else:
prob = torch.maximum(
target_probs - draft_probs,
torch.tensor(0.0, pin_memory=True).to(device, non_blocking=True),
)
q_values = q[token_to_batch] # [num_tokens, vocab_size]
epsilon = 1e-10
q_values_safe = torch.where(q_values == 0, epsilon, q_values)
q_values_safe = torch.where(torch.isinf(q_values), epsilon, q_values_safe)
prob_over_q = prob / q_values_safe
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
prob_over_q = torch.where((q_values == 0) | torch.isinf(q_values), -1e10,
prob_over_q)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
recovered_ids = torch.argmax(prob_over_q, dim=1)
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) ### What this PR does / why we need it? - Vetorize the loop (but change not output) in some rejectsampler functions include: `expand_pytorch`, `sample_recovered_tokens_pytorch`, `rejection_random_sample_pytorch`, `sample_recovered_tokens`. - Remove synchronize-launch torchnpu operator in them to accelerate sampling + MTP postprocess. ### Does this PR introduce _any_ user-facing change? - No ### How was this patch tested? - We tested this change with the serve&bench command: ``` ===== serve ===== vllm serve $LOCAL_CKPT_DIR \ --host 0.0.0.0 \ --port 8000 \ --data-parallel-size 4 \ --data-parallel-size-local 2 \ --data-parallel-address $MASTER_NODE_IP \ --data-parallel-start-rank $((2*VC_TASK_INDEX)) \ --data-parallel-rpc-port 13387 \ --tensor-parallel-size 8 \ --seed 1024 \ --enable-expert-parallel \ --served-model-name $NAME \ --max-model-len 4096 \ --max-num-seqs 16 \ --trust-remote-code \ --gpu-memory-utilization 0.90 \ $headless \ --speculative_config '{"method": "deepseek_mtp", "num_speculative_tokens": 1}' \ --additional-config '{"ascend_scheduler_config":{"enabled":false, "enable_chunked_prefill":true, "chunked_prefill_enabled":true}}' ==== bench ===== vllm bench serve --model $LOCAL_CKPT_DIR --served-model-name DeepseekV3ForCausalLM \ --dataset-name spec_bench --spec-bench-output-len 2048 \ --dataset-path question.jsonl \ --top-p 1.0 --temperature 0.8 \ --ignore-eos \ --num-prompts 64 --trust-remote-code --base-url "http://0.0.0.0:8000" --request-rate 64 ``` - In this case, our rj optimization can reduce TPOT from 84.94ms to 64.61ms, about 23% gain. ## before <img width="1068" height="830" alt="image" src="https://github.com/user-attachments/assets/278ac878-b49d-4588-b87c-316ca4d537f5" /> ## after <img width="781" height="756" alt="image" src="https://github.com/user-attachments/assets/0c6d37ad-ed77-40b3-a1be-4933c468365c" /> - vLLM version: v0.12.0 - vLLM main: https://github.com/vllm-project/vllm/commit/ad32e3e19ccf0526cb6744a5fed09a138a5fb2f9 --------- Signed-off-by: ZongYuan Zhan <zhanzy178@gmail.com> Co-authored-by: Yizhou <136800916+yiz-liu@users.noreply.github.com>
2025-12-29 14:10:39 +08:00
output_token_ids[:] = recovered_ids
Spec decode support for V1 Engine (#874) <!-- Thanks for sending a pull request! BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html --> ### What this PR does / why we need it? <!-- - Please clarify what changes you are proposing. The purpose of this section is to outline the changes and how this PR fixes the issue. If possible, please consider writing useful notes for better and faster reviews in your PR. - Please clarify why the changes are needed. For instance, the use case and bug description. - Fixes # --> Make spec decode support for V1 Engine - Currently, Ascend does not support the triton kernel. PyTorch is used to rewrite the `rejection_sampler.py` triton kernel. However, PyTorch is not as good as Triton. Therefore, ascend c is used to implement the function in the future. - Currently, spec decode supports only the ngram algorithm. The eagle algorithm needs to be further adapted. ### Does this PR introduce _any_ user-facing change? <!-- Note that it means *any* user-facing change including all aspects such as API, interface or other behavior changes. Documentation-only updates are not considered user-facing changes. --> Not change user facing. ### How was this patch tested? <!-- CI passed with new added/existing test. If it was tested in a way different from regular unit tests, please clarify how you tested step by step, ideally copy and paste-able, so that other reviewers can test and check, and descendants can verify in the future. If tests were not added, please describe why they were not added and/or why it was difficult to add. --> test by `tests/singlecard/spec_decode/e2e/test_v1_spec_decode.py` and `tests/sample/test_rejection_sampler.py`, test base function of rejection sampler and e2e function of spec decode. Signed-off-by: ponix-j <657511300@qq.com>
2025-05-23 14:25:46 +08:00
@triton.jit(do_not_specialize=["max_spec_len"])
def bonus_renew_1(
bonus_token_ids_ptr,
position,
output_token_ids_ptr,
):
bonus_token_id = tl.load(bonus_token_ids_ptr + position)
tl.store(output_token_ids_ptr + position * 2 + 1, bonus_token_id)
@triton.jit(do_not_specialize=["max_spec_len"])
def rejection_greedy_sample_spec_len_1_triton(
output_token_ids_ptr, # [batch_size, 2]
draft_token_ids_ptr, # [num_tokens]
target_argmax_ptr, # [num_tokens]
bonus_token_ids_ptr,
vec_len,
BLOCK_SIZE: tl.constexpr,
):
block_idx = tl.program_id(0)
offset = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offset < vec_len
draft_token_id = tl.load(draft_token_ids_ptr + offset, mask)
target_argmax_id = tl.load(target_argmax_ptr + offset, mask)
tl.store(output_token_ids_ptr + offset * 2, target_argmax_id, mask)
for pos in tl.range(0, BLOCK_SIZE):
draft_token_id1 = tl.get_element(draft_token_id, (pos, ))
target_argmax1 = tl.get_element(target_argmax_id, (pos, ))
position = block_idx * BLOCK_SIZE + pos
if draft_token_id1 == target_argmax1:
bonus_renew_1(
bonus_token_ids_ptr,
position,
output_token_ids_ptr,
)
@triton.jit(do_not_specialize=["max_spec_len"])
def bonus_renew(
bonus_token_ids_ptr,
position,
output_token_ids_ptr,
max_spec_len,
num_tokens1,
):
bonus_token_id = tl.load(bonus_token_ids_ptr + position)
tl.store(
output_token_ids_ptr + position * (max_spec_len + 1) + num_tokens1,
bonus_token_id)
@triton.jit(do_not_specialize=["max_spec_len"])
def rejection_greedy_sample_triton(
output_token_ids_ptr, # [batch_size, max_spec_len + 1]
cu_num_draft_tokens_ptr, # [batch_size]
draft_token_ids_ptr, # [num_tokens]
target_argmax_ptr, # [num_tokens]
bonus_token_ids_ptr, # [batch_size]
is_greedy_ptr, # [batch_size] or None
vec_len,
max_spec_len,
BLOCK_SIZE: tl.constexpr,
):
block_idx = tl.program_id(0)
offset = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offset < vec_len
if is_greedy_ptr is None:
is_greedy_mask = mask
else:
is_greedy = tl.load(is_greedy_ptr + offset, mask=mask, other=0)
is_greedy_mask = mask & (is_greedy != 0)
start_idx = tl.where(
offset == 0, 0,
tl.load(cu_num_draft_tokens_ptr + offset - 1, is_greedy_mask))
end_idx = tl.load(cu_num_draft_tokens_ptr + offset, is_greedy_mask)
num_draft_tokens = end_idx - start_idx
for pos in tl.range(0, BLOCK_SIZE):
num_tokens1 = tl.get_element(num_draft_tokens, (pos, ))
rejected = False
start_idx1 = tl.get_element(start_idx, (pos, ))
is_greedy_mask1 = tl.get_element(is_greedy_mask, (pos, ))
position = block_idx * BLOCK_SIZE + pos
for i in range(num_tokens1):
if not rejected:
draft_token_id = tl.load(draft_token_ids_ptr + start_idx1 + i)
target_argmax_id = tl.load(target_argmax_ptr + start_idx1 + i)
tl.store(
output_token_ids_ptr + position * (max_spec_len + 1) + i,
target_argmax_id,
)
if draft_token_id != target_argmax_id:
# Reject.
rejected = True
if not rejected and is_greedy_mask1:
bonus_renew(
bonus_token_ids_ptr,
position,
output_token_ids_ptr,
max_spec_len,
num_tokens1,
)
@triton.jit(do_not_specialize=["max_spec_len"])
def rejection_random_sample_kernel(
output_token_ids_ptr, # [batch_size, max_spec_len + 1]
cu_num_draft_tokens_ptr, # [batch_size]
draft_token_ids_ptr, # [num_tokens]
draft_probs_ptr, # [num_tokens, vocab_size] or None
target_probs_ptr, # [num_tokens, vocab_size]
bonus_token_ids_ptr, # [batch_size]
recovered_token_ids_ptr, # [num_tokens]
uniform_probs_ptr, # [num_tokens]
is_greedy_ptr, # [batch_size]
max_spec_len,
vocab_size,
NO_DRAFT_PROBS: tl.constexpr,
):
req_idx = tl.program_id(0)
is_greedy = tl.load(is_greedy_ptr + req_idx)
if is_greedy:
# Early exost for greedy sampling requests
return
start_idx = 0 if req_idx == 0 else tl.load(cu_num_draft_tokens_ptr +
req_idx - 1)
end_idx = tl.load(cu_num_draft_tokens_ptr + req_idx)
num_draft_tokens = end_idx - start_idx
rejected = False
for pos in range(num_draft_tokens):
if not rejected:
draft_token_id = tl.load(draft_token_ids_ptr + start_idx + pos)
if NO_DRAFT_PROBS:
draft_prob = 1
else:
draft_prob = tl.load(draft_probs_ptr +
(start_idx + pos) * vocab_size +
draft_token_id)
target_prob = tl.load(target_probs_ptr +
(start_idx + pos) * vocab_size +
draft_token_id)
uniform_prob = tl.load(uniform_probs_ptr + start_idx + pos)
if draft_prob > 0 and target_prob / draft_prob >= uniform_prob:
# Accept
token_id = draft_token_id
else:
# Reject. Use recovered token
rejected = True
token_id = tl.load(recovered_token_ids_ptr + start_idx + pos)
tl.store(output_token_ids_ptr + req_idx * (max_spec_len + 1) + pos,
token_id)
if not rejected:
# If all tokens are accepted, append the bonus token
bonus_token_id = tl.load(bonus_token_ids_ptr + req_idx)
tl.store(
output_token_ids_ptr + req_idx * (max_spec_len + 1) +
num_draft_tokens,
bonus_token_id,
)
@triton.jit(do_not_specialize=["replace_from", "replace_to"])
def expand_kernel(
output_ptr, # [num_tokens]
input_ptr, # [batch_size]
cu_num_tokens_ptr, # [batch_size]
replace_from,
replace_to,
vec_len,
MAX_NUM_TOKENS: tl.constexpr,
BLOCK_SIZE: tl.constexpr,
):
req_idx = tl.program_id(0)
offset = req_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
len_mask = offset < vec_len
start_idx = tl.where(offset == 0, 0,
tl.load(cu_num_tokens_ptr + offset - 1, len_mask))
end_idx = tl.load(cu_num_tokens_ptr + offset, len_mask)
num_tokens = end_idx - start_idx
src_val = tl.load(input_ptr + offset, len_mask)
src_val = tl.where(src_val == replace_from, replace_to, src_val)
for i in tl.range(0, BLOCK_SIZE):
num_tokens1 = tl.get_element(num_tokens, (i, ))
start_idx1 = tl.get_element(start_idx, (i, ))
src_val1 = tl.get_element(src_val, (i, ))
offset1 = tl.arange(0, MAX_NUM_TOKENS)
tl.store(output_ptr + start_idx1 + offset1,
src_val1,
mask=offset1 < num_tokens1)
@triton.jit
def sample_recovered_tokens_kernel(
output_token_ids_ptr, # [num_tokens]
cu_num_draft_tokens_ptr, # [batch_size]
draft_token_ids_ptr, # [num_tokens]
draft_probs_ptr, # [num_tokens, vocab_size] or None
target_probs_ptr, # [num_tokens, vocab_size]
q_ptr, # [batch_size, vocab_size]
vocab_size,
PADDED_VOCAB_SIZE: tl.constexpr,
NO_DRAFT_PROBS: tl.constexpr,
SUB_BLOCK: tl.constexpr,
):
req_idx = tl.program_id(0)
start_idx = 0 if req_idx == 0 else tl.load(cu_num_draft_tokens_ptr +
req_idx - 1)
end_idx = tl.load(cu_num_draft_tokens_ptr + req_idx)
num_draft_tokens = end_idx - start_idx
# Early exit for out-of-range positions.
pos = tl.program_id(1)
if pos >= num_draft_tokens:
return
loop = (vocab_size + SUB_BLOCK - 1) // SUB_BLOCK
global_recovered_id = -1
global_max_p = -1.0
if NO_DRAFT_PROBS:
draft_token_id = tl.load(draft_token_ids_ptr + start_idx + pos)
orig_prob = tl.load(target_probs_ptr + (start_idx + pos) * vocab_size +
draft_token_id)
# Temporarily zero out the probability of the draft token.
# This is essentially the same as target_prob - draft_prob, except that
# n-gram does not have draft_prob. We regard it as 1.
tl.store(
target_probs_ptr + (start_idx + pos) * vocab_size + draft_token_id,
0)
for loop_i in range(loop):
vocab_start = loop_i * SUB_BLOCK
vocab_offset = vocab_start + tl.arange(0, SUB_BLOCK)
prob = tl.load(target_probs_ptr + (start_idx + pos) * vocab_size +
vocab_offset,
mask=vocab_offset < vocab_size,
other=0)
q = tl.load(q_ptr + req_idx * vocab_size + vocab_offset,
mask=vocab_offset < vocab_size,
other=float("-inf"))
new_p = prob / q
recovered_id = tl.argmax(new_p, axis=-1)
max_p = tl.get_element(new_p, (recovered_id, ))
if max_p > global_max_p:
global_max_p = max_p
global_recovered_id = vocab_start + recovered_id
else:
for loop_i in range(loop):
vocab_start = loop_i * SUB_BLOCK
vocab_offset = vocab_start + tl.arange(0, SUB_BLOCK)
draft_prob = tl.load(draft_probs_ptr +
(start_idx + pos) * vocab_size + vocab_offset,
mask=vocab_offset < vocab_size,
other=0)
target_prob = tl.load(target_probs_ptr +
(start_idx + pos) * vocab_size +
vocab_offset,
mask=vocab_offset < vocab_size,
other=0)
prob = tl.maximum(target_prob - draft_prob, 0)
# NOTE(woosuk): We don't need `prob = prob / tl.sum(prob)` here because
# `tl.argmax` will select the maximum value.
q = tl.load(q_ptr + req_idx * vocab_size + vocab_offset,
mask=vocab_offset < vocab_size,
other=float("-inf"))
new_p = prob / q
recovered_id = tl.argmax(new_p, axis=-1)
max_p = tl.get_element(new_p, (recovered_id, ))
if max_p > global_max_p:
global_max_p = max_p
global_recovered_id = vocab_start + recovered_id
tl.store(output_token_ids_ptr + start_idx + pos, global_recovered_id)
if NO_DRAFT_PROBS:
# Restore the original probability.
tl.store(
target_probs_ptr + (start_idx + pos) * vocab_size + draft_token_id,
orig_prob)