chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

use flashinfer.sampling (#18696)

Browse Source
This commit is contained in:
pansicheng
2026-02-26 10:02:38 +08:00
committed by GitHub
parent 2739d7df62
commit 2ad475b4ed
8 changed files with 14 additions and 431 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/layers/sampler.py
View File

@@ -19,13 +19,14 @@ from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils.common import crash_on_warnings, get_bool_env_var, is_cuda, is_npu
if is_cuda():
from sgl_kernel import (
from flashinfer.sampling import (
min_p_sampling_from_probs,
top_k_renorm_prob,
top_k_top_p_sampling_from_probs,
)
from sgl_kernel import (
top_k_renorm_prob,
top_p_renorm_prob,
)
if is_npu():
import torch_npu

1
sgl-kernel/CMakeLists.txt
View File

@@ -329,7 +329,6 @@ set(SOURCES
"${repo-flashinfer_SOURCE_DIR}/csrc/norm.cu"
"${repo-flashinfer_SOURCE_DIR}/csrc/renorm.cu"
"${repo-flashinfer_SOURCE_DIR}/csrc/sampling.cu"
"${repo-flash-attention_SOURCE_DIR}/csrc/flash_attn/src/flash_fwd_sparse_hdim128_bf16_causal_sm80.cu"
"${repo-flash-attention_SOURCE_DIR}/csrc/flash_attn/src/flash_fwd_sparse_hdim128_bf16_sm80.cu"

5
sgl-kernel/benchmark/bench_top_k_top_p_sampling.py
View File

@@ -1,6 +1,7 @@
import itertools
import os
import flashinfer.sampling
import sgl_kernel
import torch
import triton
@@ -69,7 +70,7 @@ def calculate_diff(batch_size, vocab_size, p):
torch_samples = torch_top_k_top_p_joint_sampling_from_probs(
normalized_prob, top_k_tensor, top_p_tensor
)
sglang_samples = sgl_kernel.top_k_top_p_sampling_from_probs(
sglang_samples = flashinfer.sampling.top_k_top_p_sampling_from_probs(
normalized_prob, top_k_tensor, top_p_tensor, filter_apply_order="joint"
)
@@ -120,7 +121,7 @@ def benchmark_sampling(batch_size, vocab_size, p, provider):
normalized_prob.clone(), top_k_tensor, top_p_tensor
)
elif provider == "sglang":
fn = lambda: sgl_kernel.top_k_top_p_sampling_from_probs(
fn = lambda: flashinfer.sampling.top_k_top_p_sampling_from_probs(
normalized_prob.clone(),
top_k_tensor,
top_p_tensor,

15
sgl-kernel/csrc/common_extension.cc
View File

@@ -417,27 +417,12 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
{at::Tag::needs_fixed_stride_order});
m.impl("bmm_fp8", torch::kCUDA, &bmm_fp8);
m.def(
"min_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? maybe_min_p_arr, float "
"min_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("min_p_sampling_from_probs", torch::kCUDA, &min_p_sampling_from_probs);
m.def("top_k_renorm_probs(Tensor probs, Tensor! renorm_probs, Tensor? maybe_top_k_arr, int top_k_val) -> ()");
m.impl("top_k_renorm_probs", torch::kCUDA, &top_k_renorm_probs);
m.def("top_p_renorm_probs(Tensor probs, Tensor! renorm_probs, Tensor? maybe_top_p_arr, float top_p_val) -> ()");
m.impl("top_p_renorm_probs", torch::kCUDA, &top_p_renorm_probs);
m.def(
"top_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? "
"maybe_top_p_arr, float top_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("top_p_sampling_from_probs", torch::kCUDA, &top_p_sampling_from_probs);
m.def(
"top_k_top_p_sampling_from_probs(Tensor probs, Tensor output, Tensor? maybe_indices, Tensor? maybe_top_k_arr, "
"float top_k_val, Tensor? maybe_top_p_arr, float top_p_val, bool deterministic, Generator? gen) -> ()");
m.impl("top_k_top_p_sampling_from_probs", torch::kCUDA, &top_k_top_p_sampling_from_probs);
m.def("top_k_mask_logits(Tensor logits, Tensor mask_logits, Tensor? maybe_top_k_arr, int top_k_val) -> ()");
m.impl("top_k_mask_logits", torch::kCUDA, &top_k_mask_logits);

29
sgl-kernel/include/sgl_kernel_ops.h
View File

@@ -683,41 +683,12 @@ void store_kv_cache(at::Tensor k_cache, at::Tensor v_cache, at::Tensor out_loc,
/*
* From FlashInfer
*/
void min_p_sampling_from_probs(
at::Tensor probs,
at::Tensor output,
std::optional<at::Tensor> maybe_indices,
std::optional<at::Tensor> maybe_min_p_arr,
double min_p_val,
bool deterministic,
std::optional<at::Generator> gen);
void top_k_renorm_probs(
at::Tensor probs, at::Tensor renorm_probs, std::optional<at::Tensor> maybe_top_k_arr, int64_t top_k_val);
void top_p_renorm_probs(
at::Tensor probs, at::Tensor renorm_probs, std::optional<at::Tensor> maybe_top_p_arr, double top_p_val);
void top_k_top_p_sampling_from_probs(
at::Tensor probs,
at::Tensor output,
std::optional<at::Tensor> maybe_indices,
std::optional<at::Tensor> maybe_top_k_arr,
double top_k_val,
std::optional<at::Tensor> maybe_top_p_arr,
double top_p_val,
bool deterministic,
std::optional<at::Generator> gen);
void top_p_sampling_from_probs(
at::Tensor probs,
at::Tensor output,
std::optional<at::Tensor> maybe_indices,
std::optional<at::Tensor> maybe_top_p_arr,
double top_p_val,
bool deterministic,
std::optional<at::Generator> gen);
void top_k_mask_logits(
at::Tensor logits, at::Tensor mask_logits, std::optional<at::Tensor> maybe_top_k_arr, int64_t top_k_val);

4
sgl-kernel/python/sgl_kernel/__init__.py
View File

@@ -101,13 +101,9 @@ from sgl_kernel.quantization import (
ggml_mul_mat_vec_a8,
)
from sgl_kernel.sampling import (
min_p_sampling_from_probs,
top_k_mask_logits,
top_k_renorm_prob,
top_k_top_p_sampling_from_logits,
top_k_top_p_sampling_from_probs,
top_p_renorm_prob,
top_p_sampling_from_probs,
)
from sgl_kernel.speculative import (
build_tree_kernel_efficient,

371
sgl-kernel/python/sgl_kernel/sampling.py
View File

@@ -102,289 +102,6 @@ def top_p_renorm_probs(
top_p_renorm_prob = top_p_renorm_probs
def _top_p_sampling_from_probs_internal(
probs: torch.Tensor,
indices: Optional[torch.Tensor],
maybe_top_p_arr: Optional[torch.Tensor],
top_p_val: float,
deterministic: bool,
generator: Optional[torch.Generator],
) -> torch.Tensor:
with probs.device as device:
probs = probs.float()
maybe_top_p_arr = (
maybe_top_p_arr.float() if maybe_top_p_arr is not None else None
)
samples = torch.empty(probs.size(0), dtype=torch.int32, device=device)
torch.ops.sgl_kernel.top_p_sampling_from_probs.default(
probs,
samples,
indices,
maybe_top_p_arr,
top_p_val,
deterministic,
generator,
)
return samples
def top_p_sampling_from_probs(
probs: torch.Tensor,
top_p: Union[torch.Tensor, float],
indices: Optional[torch.Tensor] = None,
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for top-p sampling (nucleus sampling) from probabilities,
this operator implements GPU-based rejection sampling without explicit sorting.
Check the `blog post <https://flashinfer.ai/2025/03/10/sampling.html>`_ for more details.
The multiple rounds of rejection sampling are implemented in a single CUDA kernel,
which is more efficient than the naive implementation that launches a series of kernels.
Parameters
----------
probs: torch.Tensor
Probabilities for sampling. When indices is not provided, shape should be ``(batch_size, num_classes)``
and the i-th output will be sampled from the i-th row of probabilities. When indices is provided,
shape should be ``(unique_batch_size, num_classes)`` where unique_batch_size is the number of unique
probability distributions.
top_p: Union[torch.Tensor, float]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for top-p sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
indices: Optional[torch.Tensor]
Optional indices tensor of shape ``(batch_size,)`` that maps each output to a row in probs.
For example, if indices[i] = j, then the i-th output will be sampled from probs[j].
This allows reusing the same probability distribution for multiple outputs.
If indices is not provided, the i-th output will be sampled from the i-th row of probs.
deterministic: bool
Whether to use deterministic kernel implementation, default is ``True``.
generator: Optional[torch.Generator]
A random number generator for the operation.
check_nan: bool
Whether to check nan in :attr:`probs`, default is ``False``.
Returns
-------
samples: torch.Tensor
Sampled categories, shape ``(batch_size,)``.
Note
----
This function expects float32 inputs, and the output is int32.
"""
if check_nan:
if torch.any(torch.isnan(probs)):
raise ValueError("Input probs contains NaN.")
return _top_p_sampling_from_probs_internal(
probs, indices, *_to_tensor_scalar_tuple(top_p), deterministic, generator
)
def _top_k_top_p_sampling_from_probs_internal(
probs: torch.Tensor,
indices: Optional[torch.Tensor],
maybe_top_k_arr: Optional[torch.Tensor],
top_k_val: int,
maybe_top_p_arr: Optional[torch.Tensor],
top_p_val: float,
deterministic: bool,
generator: Optional[torch.Generator],
) -> torch.Tensor:
with probs.device as device:
probs = probs.float()
maybe_top_k_arr = maybe_top_k_arr.int() if maybe_top_k_arr is not None else None
maybe_top_p_arr = (
maybe_top_p_arr.float() if maybe_top_p_arr is not None else None
)
samples = torch.empty(probs.size(0), dtype=torch.int32, device=device)
torch.ops.sgl_kernel.top_k_top_p_sampling_from_probs.default(
probs,
samples,
indices,
maybe_top_k_arr,
top_k_val,
maybe_top_p_arr,
top_p_val,
deterministic,
generator,
)
return samples
def top_k_top_p_sampling_from_probs(
probs: torch.Tensor,
top_k: Union[torch.Tensor, int],
top_p: Union[torch.Tensor, float],
indices: Optional[torch.Tensor] = None,
filter_apply_order: str = "top_k_first",
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for top-k and top-p sampling from probabilities,
this operator implements GPU-based rejection sampling without explicit sorting.
Check the `blog post <https://flashinfer.ai/2025/03/10/sampling.html>`_ for more details.
The multiple rounds of rejection sampling are implemented in a single CUDA kernel,
which is more efficient than the naive implementation that launches a series of kernels.
Parameters
----------
probs: torch.Tensor
Probabilities for sampling. When indices is not provided, shape should be ``(batch_size, num_classes)``
and the i-th output will be sampled from the i-th row of probabilities. When indices is provided,
shape should be ``(unique_batch_size, num_classes)`` where unique_batch_size is the number of unique
probability distributions.
top_k: Union[torch.Tensor, int]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for top-k sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
top_p: Union[torch.Tensor, float]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for top-p sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
indices: Optional[torch.Tensor]
Optional indices tensor of shape ``(batch_size,)`` that maps each output to a row in probs.
For example, if indices[i] = j, then the i-th output will be sampled from probs[j].
This allows reusing the same probability distribution for multiple outputs.
If indices is not provided, the i-th output will be sampled from the i-th row of probs.
filter_apply_order: str
The order of applying top-k and top-p sampling, should be either ``"top_k_first"`` or ``"joint"``.
If ``"top_k_first"``, we first apply top-k filter, then apply top-p sampling on the top-k results.
If ``"joint"``, we apply top-k and top-p filter simultaneously in each round. Default is ``"top_k_first"``.
deterministic: bool
Whether to use deterministic kernel implementation, default is ``True``.
generator: Optional[torch.Generator]
A random number generator for the operation.
check_nan: bool
Whether to check nan in :attr:`probs`, default is ``False``.
Returns
-------
samples: torch.Tensor
Sampled categories, shape ``(batch_size,)``.
Note
----
This function expects float32 inputs, and the output is int32.
"""
if filter_apply_order == "top_k_first":
renorm_probs = top_k_renorm_probs(probs, top_k)
return top_p_sampling_from_probs(
renorm_probs,
top_p,
indices,
deterministic,
check_nan=check_nan,
generator=generator,
)
elif filter_apply_order == "joint":
if check_nan:
if torch.any(torch.isnan(probs)):
raise ValueError("Input probs contains NaN.")
return _top_k_top_p_sampling_from_probs_internal(
probs,
indices,
*_to_tensor_scalar_tuple(top_k),
*_to_tensor_scalar_tuple(top_p),
deterministic,
generator,
)
else:
raise ValueError(f"Invalid filter_apply_order: {filter_apply_order}")
def _min_p_sampling_from_probs_internal(
probs: torch.Tensor,
indices: Optional[torch.Tensor],
maybe_min_p_arr: Optional[torch.Tensor],
min_p_val: float,
deterministic: bool,
generator: Optional[torch.Generator],
) -> torch.Tensor:
with probs.device as device:
probs = probs.float()
maybe_min_p_arr = (
maybe_min_p_arr.float() if maybe_min_p_arr is not None else None
)
samples = torch.empty(probs.size(0), dtype=torch.int32, device=device)
torch.ops.sgl_kernel.min_p_sampling_from_probs.default(
probs,
samples,
indices,
maybe_min_p_arr,
min_p_val,
deterministic,
generator,
)
return samples
def min_p_sampling_from_probs(
probs: torch.Tensor,
min_p: Union[torch.Tensor, float],
indices: Optional[torch.Tensor] = None,
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for `min_p sampling <https://arxiv.org/abs/2407.01082>`_ from probabilities,
this operator implements GPU-based rejection sampling without explicit sorting.
Check the `blog post <https://flashinfer.ai/2025/03/10/sampling.html>`_ for more details.
The multiple rounds of rejection sampling are implemented in a single CUDA kernel,
which is more efficient than the naive implementation that launches a series of kernels.
Parameters
----------
probs: torch.Tensor
Probabilities for sampling. When indices is not provided, shape should be ``(batch_size, num_classes)``
and the i-th output will be sampled from the i-th row of probabilities. When indices is provided,
shape should be ``(unique_batch_size, num_classes)`` where unique_batch_size is the number of unique
probability distributions.
min_p: Union[torch.Tensor, float]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for min-p sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
indices: Optional[torch.Tensor]
Optional indices tensor of shape ``(batch_size,)`` that maps each output to a row in probs.
For example, if indices[i] = j, then the i-th output will be sampled from probs[j].
This allows reusing the same probability distribution for multiple outputs.
If indices is not provided, the i-th output will be sampled from the i-th row of probs.
deterministic: bool
Whether to use deterministic kernel implementation, default is ``True``.
generator: Optional[torch.Generator]
A random number generator for the operation.
check_nan: bool
Whether to check nan in :attr:`probs`, default is ``False``.
Returns
-------
samples: torch.Tensor
Sampled categories, shape ``(batch_size,)``.
Note
----
This function expects float32 inputs, and the output is int32.
"""
if check_nan:
if torch.any(torch.isnan(probs)):
raise ValueError("Input probs contains NaN.")
return _min_p_sampling_from_probs_internal(
probs, indices, *_to_tensor_scalar_tuple(min_p), deterministic, generator
)
def _top_k_mask_logits_internal(
logits: torch.Tensor,
maybe_top_k_arr: Optional[torch.Tensor],
@@ -453,91 +170,3 @@ def top_k_mask_logits(
top_k_renorm_probs
"""
return _top_k_mask_logits_internal(logits, *_to_tensor_scalar_tuple(top_k))
def top_k_top_p_sampling_from_logits(
logits: torch.Tensor,
top_k: Union[torch.Tensor, int],
top_p: Union[torch.Tensor, float],
indices: Optional[torch.Tensor] = None,
filter_apply_order: str = "top_k_first",
deterministic: bool = True,
generator: Optional[torch.Generator] = None,
check_nan: bool = False,
) -> torch.Tensor:
r"""Adapt from https://github.com/flashinfer-ai/flashinfer/flashinfer/sampling.py
Fused GPU kernel for top-k and top-p sampling from probabilities,
this operator implements GPU-based rejection sampling without explicit sorting.
Check the `blog post <https://flashinfer.ai/2025/03/10/sampling.html>`_ for more details.
The multiple rounds of rejection sampling are implemented in a single CUDA kernel,
which is more efficient than the naive implementation that launches a series of kernels.
Parameters
----------
logits: torch.Tensor
Pre-softmax logits for sampling. When indices is not provided, shape should be ``(batch_size, num_classes)``
and the i-th output will be sampled from the i-th row of logits. When indices is provided,
shape should be ``(unique_batch_size, num_classes)`` where unique_batch_size is the number of unique
probability distributions.
top_k: Union[torch.Tensor, int]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for top-k sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
top_p: Union[torch.Tensor, float]
Either a scalar or a tensor of shape ``(batch_size,)``, representing the threshold for top-p sampling.
If a scalar, the same threshold is used for all requests.
If a tensor, each request has its own threshold.
indices: Optional[torch.Tensor]
Optional indices tensor of shape ``(batch_size,)`` that maps each output to a row in probs.
For example, if indices[i] = j, then the i-th output will be sampled from probs[j].
This allows reusing the same probability distribution for multiple outputs.
If indices is not provided, the i-th output will be sampled from the i-th row of probs.
filter_apply_order: str
The order of applying top-k and top-p sampling, should be either ``"top_k_first"`` or ``"joint"``.
If ``"top_k_first"``, we first apply top-k filter, then apply top-p sampling on the top-k results.
If ``"joint"``, we apply top-k and top-p filter simultaneously in each round. Default is ``"top_k_first"``.
deterministic: bool
Whether to use deterministic kernel implementation, default is ``True``.
generator: Optional[torch.Generator]
A random number generator for the operation.
check_nan: bool
Whether to check nan in :attr:`probs`, default is ``False``.
Returns
-------
samples: torch.Tensor
Sampled categories, shape ``(batch_size,)``.
Note
----
This function expects float32 inputs, and the output is int32.
"""
if filter_apply_order == "top_k_first":
masked_logits = top_k_mask_logits(logits, top_k)
probs = torch.softmax(masked_logits, dim=-1)
return top_p_sampling_from_probs(
probs,
top_p,
indices,
deterministic,
check_nan=check_nan,
generator=generator,
)
elif filter_apply_order == "joint":
probs = torch.softmax(logits, dim=-1)
if check_nan:
if torch.any(torch.isnan(probs)):
raise ValueError("Input probs contains NaN.")
return _top_k_top_p_sampling_from_probs_internal(
probs,
indices,
*_to_tensor_scalar_tuple(top_k),
*_to_tensor_scalar_tuple(top_p),
deterministic,
generator,
)
else:
raise ValueError(f"Invalid filter_apply_order: {filter_apply_order}")

13
sgl-kernel/tests/test_sampling.py
View File

@@ -1,5 +1,6 @@
# Adapted from https://github.com/flashinfer-ai/flashinfer/blob/93e1a2634e22355b0856246b032b285ad1d1da6b/tests/test_sampling.py
import flashinfer.sampling
import pytest
import sgl_kernel
import torch
@@ -16,10 +17,10 @@ def test_top_k_top_p_sampling_from_probs_logits_top_k_first_alignment(
logits = torch.randn(batch_size, vocab_size, device="cuda:0") * 5
generator_logits = torch.Generator("cuda:0")
generator_probs = generator_logits.clone_state()
samples = sgl_kernel.sampling.top_k_top_p_sampling_from_logits(
samples = flashinfer.sampling.top_k_top_p_sampling_from_logits(
logits, k, p, filter_apply_order="top_k_first", generator=generator_logits
)
samples_ref = sgl_kernel.sampling.top_k_top_p_sampling_from_probs(
samples_ref = flashinfer.sampling.top_k_top_p_sampling_from_probs(
torch.softmax(logits, dim=-1),
k,
p,
@@ -40,10 +41,10 @@ def test_top_k_top_p_sampling_from_probs_logits_joint_alignment(
logits = torch.randn(batch_size, vocab_size, device="cuda:0") * 5
generator_logits = torch.Generator("cuda:0")
generator_probs = generator_logits.clone_state()
samples = sgl_kernel.sampling.top_k_top_p_sampling_from_logits(
samples = flashinfer.sampling.top_k_top_p_sampling_from_logits(
logits, k, p, filter_apply_order="joint", generator=generator_logits
)
samples_ref = sgl_kernel.sampling.top_k_top_p_sampling_from_probs(
samples_ref = flashinfer.sampling.top_k_top_p_sampling_from_probs(
torch.softmax(logits, dim=-1),
k,
p,
@@ -83,7 +84,7 @@ def test_top_k_top_p_joint_sampling_from_probs(batch_size, vocab_size, p):
num_trails = 1000
for _ in range(num_trails):
samples = sgl_kernel.top_k_top_p_sampling_from_probs(
samples = flashinfer.sampling.top_k_top_p_sampling_from_probs(
normalized_prob,
top_k_tensor,
top_p_tensor,
@@ -167,7 +168,7 @@ def test_min_p_sampling(batch_size, vocab_size, p):
num_trails = 1000
for _ in range(num_trails):
samples = sgl_kernel.min_p_sampling_from_probs(
samples = flashinfer.sampling.min_p_sampling_from_probs(
normalized_prob,
min_p_tensor,
)