EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

feat: implement high-performance Triton kernels for rejection sampling: optimization for rejection_random_sample_kernel (#5259)

Browse Source 
### What this PR does / why we need it?

This PR introduces optimized Triton implementations for the
rejection_random_sample_kernel delivering superior performance compared
to the existing Triton implementations. The new Triton kernels maintain
full functional accuracy while delivering significant performance
improvements across various batch sizes and MTP configurations.

### Does this PR introduce _any_ user-facing change?

Yes, this PR modifies rejection_sampler.py to use optimized Triton
kernels:
rejection_random_sample_kernel is modified and optimized

### How was this patch tested?
performance benchmark results:
<html xmlns:v="urn:schemas-microsoft-com:vml"
xmlns:o="urn:schemas-microsoft-com:office:office"
xmlns:x="urn:schemas-microsoft-com:office:excel"
xmlns="http://www.w3.org/TR/REC-html40">
<head>

<meta name=Generator content="Microsoft Excel">
<!--[if !mso]>
</head>
<body>
<!--StartFragment-->

Batch Size | MTP | origin implementation(us) | optimized version(us)
-- | -- | -- | --
1 | 1 | 2.934 | 3.64
8 | 1 | 4.467 | 4
32 | 1 | 6.98 | 4.54
64 | 1 | 11.087 | 6.42
128 | 1 | 13.414 | 7.84
256 | 1 | 19.66 | 8.487
512 | 1 | 39.908 | 11.62
1024 | 1 | 81.781 | 18.16
2048 | 1 | 137.923 | 32.934
1 | 2 | 3.4 | 4.02
8 | 2 | 3.74 | 4.24
32 | 2 | 6.373 | 7.394
64 | 2 | 9.747 | 6.46
128 | 2 | 12.98 | 7.76
256 | 2 | 20.834 | 9.787
512 | 2 | 39.314 | 13.56
1024 | 2 | 83.135 | 22.387
2048 | 2 | 157.563 | 40.607


<!--EndFragment-->
</body>

</html>


- vLLM version: release/v0.13.0
- vLLM main:
ad32e3e19c

Signed-off-by: 1024daniel <xxltju324@gmail.com>
This commit is contained in:
daniel
2026-01-05 16:03:02 +08:00
committed by GitHub
parent 91bf524364
commit 8ffe3f5d78
3 changed files with 195 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

95
tests/e2e/nightly/ops/triton/test_rejection_sample.py Normal file
View File

@@ -0,0 +1,95 @@
import pytest
import torch
from vllm.v1.sample.rejection_sampler import \
rejection_random_sample_kernel as original_rejection_random_sample_kernel
from vllm_ascend.ops.triton.reject_sample import (
cal_grid_and_block_size, rejection_random_sample_kernel)
from vllm_ascend.ops.triton.triton_utils import init_device_properties_triton
@pytest.fixture(scope="function", autouse=True)
def setup_device_properties():
init_device_properties_triton()
yield
@pytest.mark.parametrize("max_spec_len", [1, 2, 3])
@pytest.mark.parametrize("vocab_size", [151_936])
@pytest.mark.parametrize("batch_size", [1, 8, 32, 64, 128, 256, 512, 1024])
@torch.inference_mode()
def test_rejection_random_sample(max_spec_len, vocab_size, batch_size):
device = 'npu'
torch.manual_seed(0)
draft_probs = torch.rand(batch_size * max_spec_len,
vocab_size,
dtype=torch.float32,
device=device)
target_probs = torch.rand(batch_size * max_spec_len,
vocab_size,
dtype=torch.float32,
device=device)
bonus_token_ids = torch.randint(low=0,
high=vocab_size,
size=(batch_size, 1),
dtype=torch.int64,
device=device)
draft_token_ids = torch.randint(low=0,
high=vocab_size,
size=(batch_size * max_spec_len, ),
dtype=torch.int64,
device=device)
output_token_ids = torch.empty((batch_size, max_spec_len + 1),
dtype=torch.int64,
device=device)
original_output_token_ids = output_token_ids.clone()
num_tokens = draft_token_ids.shape[0]
uniform_probs = torch.rand((num_tokens, ),
dtype=torch.float32,
device=device)
num_draft_tokens = [max_spec_len] * batch_size
num_draft_tokens = torch.tensor(num_draft_tokens,
dtype=torch.int32,
device=device)
cu_num_draft_tokens = torch.cumsum(num_draft_tokens,
dim=0,
dtype=torch.int32)
is_greedy_ptr = torch.full((batch_size, ),
False,
dtype=torch.bool,
device=device)
recovered_ids = torch.zeros_like(draft_token_ids,
dtype=torch.int64,
device=device)
grid, block_size = cal_grid_and_block_size(batch_size)
original_rejection_random_sample_kernel[(batch_size, )](
original_output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
bonus_token_ids,
recovered_ids,
uniform_probs,
is_greedy_ptr,
max_spec_len,
vocab_size,
NO_DRAFT_PROBS=draft_probs is None,
)
rejection_random_sample_kernel[(grid, )](output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
draft_probs,
target_probs,
bonus_token_ids,
recovered_ids,
uniform_probs,
is_greedy_ptr,
max_spec_len,
vocab_size,
batch_size,
NO_DRAFT_PROBS=draft_probs
is None,
BLOCK_SIZE=block_size)
torch.npu.synchronize()
assert torch.equal(original_output_token_ids, output_token_ids)

169
vllm_ascend/ops/triton/reject_sample.py
View File

@@ -20,6 +20,17 @@ from vllm.triton_utils import tl, triton
from vllm_ascend.ops.triton.triton_utils import get_vectorcore_num
def cal_grid_and_block_size(batch_size: int):
vectorcore_num = get_vectorcore_num()
if batch_size <= vectorcore_num:
grid = batch_size
block_size = 1
else:
grid = vectorcore_num
block_size = triton.next_power_of_2(triton.cdiv(batch_size, grid))
return grid, block_size
@triton.jit(do_not_specialize=["max_spec_len"])
def bonus_renew_1(
bonus_token_ids_ptr,
@@ -131,62 +142,72 @@ def rejection_greedy_sample_triton(
@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,
)
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,
vec_len,
NO_DRAFT_PROBS: tl.constexpr,
BLOCK_SIZE: tl.constexpr):
block_idx = tl.program_id(0)
offsets = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < vec_len
is_greedy = tl.load(is_greedy_ptr + offsets, mask, other=1)
not_greedy_mask = is_greedy == 0
start_idxs = tl.where(
offsets == 0, 0,
tl.load(cu_num_draft_tokens_ptr + offsets - 1, not_greedy_mask))
end_idxs = tl.load(cu_num_draft_tokens_ptr + offsets, not_greedy_mask)
n_num_draft_tokens = end_idxs - start_idxs
for req_i in range(BLOCK_SIZE):
not_greedy = tl.get_element(not_greedy_mask, (req_i, ))
if not_greedy:
rejected = False
start_idx = tl.get_element(start_idxs, (req_i, ))
req_idx = block_idx * BLOCK_SIZE + req_i
num_draft_tokens = tl.get_element(n_num_draft_tokens, (req_i, ))
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)
# NOTE(woosuk): While the draft probability should never be 0,
# we check it to avoid NaNs. If it happens to be 0, we reject.
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"])
@@ -311,24 +332,12 @@ def sample_recovered_tokens_kernel(
orig_prob)
def rejection_greedy_sample_with_triton(
output_token_ids,
num_draft_tokens,
cu_num_draft_tokens,
draft_token_ids,
target_argmax,
bonus_token_ids,
is_greedy,
max_spec_len,
):
def rejection_greedy_sample_with_triton(output_token_ids, num_draft_tokens,
cu_num_draft_tokens, draft_token_ids,
target_argmax, bonus_token_ids,
is_greedy, max_spec_len, grid,
block_size):
vec_len = output_token_ids.shape[0]
n = cu_num_draft_tokens.numel()
BLOCK_SIZE = 2
grid = triton.cdiv(n, BLOCK_SIZE)
vectorcore_num = get_vectorcore_num()
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 is_greedy is None:
@@ -338,7 +347,7 @@ def rejection_greedy_sample_with_triton(
target_argmax,
bonus_token_ids,
vec_len,
BLOCK_SIZE=BLOCK_SIZE,
BLOCK_SIZE=block_size,
)
else:
rejection_greedy_sample_triton[(grid, )](
@@ -350,20 +359,14 @@ def rejection_greedy_sample_with_triton(
is_greedy,
vec_len,
max_spec_len,
BLOCK_SIZE=BLOCK_SIZE,
BLOCK_SIZE=block_size,
)
def expand_triton(batch_size, expanded_x, x, cu_num_tokens, replace_from,
replace_to, max_num_tokens):
vec_len = batch_size
n = cu_num_tokens.numel()
BLOCK_SIZE = 2
grid = triton.cdiv(n, BLOCK_SIZE)
vectorcore_num = get_vectorcore_num()
if n >= vectorcore_num:
grid = vectorcore_num
BLOCK_SIZE = triton.next_power_of_2(triton.cdiv(n, grid))
grid, block_size = cal_grid_and_block_size(batch_size)
expand_kernel[(grid, )](
expanded_x,
@@ -373,5 +376,5 @@ def expand_triton(batch_size, expanded_x, x, cu_num_tokens, replace_from,
replace_to,
vec_len,
MAX_NUM_TOKENS=max_num_tokens, # To avoid recompilation.
BLOCK_SIZE=BLOCK_SIZE,
BLOCK_SIZE=block_size,
)

27
vllm_ascend/sample/rejection_sampler.py
View File

@@ -8,8 +8,9 @@ from vllm.v1.sample.rejection_sampler import (GREEDY_TEMPERATURE,
generate_uniform_probs)
from vllm_ascend.ops.triton.reject_sample import (
expand_triton, rejection_greedy_sample_with_triton,
rejection_random_sample_kernel, sample_recovered_tokens_kernel)
cal_grid_and_block_size, expand_triton,
rejection_greedy_sample_with_triton, rejection_random_sample_kernel,
sample_recovered_tokens_kernel)
from vllm_ascend.sample.sampler import apply_top_k_top_p
PLACEHOLDER_TOKEN_ID = -1
@@ -119,20 +120,18 @@ def rejection_sample(
is_greedy = None
else:
is_greedy = sampling_metadata.temperature == GREEDY_TEMPERATURE
if HAS_TRITON:
grid, block_size = cal_grid_and_block_size(batch_size)
if not sampling_metadata.all_random:
# Rejection sampling for greedy sampling requests.
target_argmax = target_probs.argmax(dim=-1)
if HAS_TRITON:
rejection_greedy_sample_with_triton(
output_token_ids,
num_draft_tokens,
cu_num_draft_tokens,
draft_token_ids,
target_argmax,
bonus_token_ids,
is_greedy,
max_spec_len,
)
rejection_greedy_sample_with_triton(output_token_ids,
num_draft_tokens,
cu_num_draft_tokens,
draft_token_ids, target_argmax,
bonus_token_ids, is_greedy,
max_spec_len, grid, block_size)
else:
if min(num_draft_tokens) == 1 and max(
num_draft_tokens) == 1 and sampling_metadata.all_greedy:
@@ -180,7 +179,7 @@ def rejection_sample(
# Rejection sampling for random sampling requests.
if HAS_TRITON:
rejection_random_sample_kernel[(batch_size, )](
rejection_random_sample_kernel[(grid, )](
output_token_ids,
cu_num_draft_tokens,
draft_token_ids,
@@ -192,7 +191,9 @@ def rejection_sample(
is_greedy,
max_spec_len,
vocab_size,
batch_size,
NO_DRAFT_PROBS=draft_probs is None,
BLOCK_SIZE=block_size,
)
else:
rejection_random_sample_pytorch(