EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

update variable naming and comments for rocm (#5299)

Browse Source
This commit is contained in:
Zhaoyi Li
2025-04-12 01:15:05 -05:00
committed by GitHub
parent 2eb55770f9
commit 3c9740d200
6 changed files with 14 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py
View File

@@ -11,8 +11,8 @@ from vllm import _custom_ops as ops
from sglang.srt.utils import is_hip
is_hip_ = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
def vllm_scaled_fp8_quant(

4
sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
View File

@@ -8,8 +8,8 @@ from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_group_quant_
from sglang.srt.utils import is_hip
is_hip_ = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
@triton.jit

4
sgl-kernel/benchmark/bench_per_token_quant_fp8.py
View File

@@ -9,8 +9,8 @@ from vllm import _custom_ops as ops
from sglang.srt.utils import is_hip
is_hip_ = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
def vllm_per_token_quant_fp8(

4
sgl-kernel/csrc/torch_extension_rocm.cc
View File

@@ -61,11 +61,15 @@ TORCH_LIBRARY_EXPAND(sgl_kernel, m) {
"moe_align_block_size(Tensor topk_ids, int num_experts, int block_size, Tensor! sorted_token_ids, Tensor! "
"experts_ids, Tensor! num_tokens_post_pad, Tensor! token_cnts_buffer, Tensor! cumsum_buffer) -> ()");
m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);
m.def(
"topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
"token_expert_indices, Tensor gating_output) -> ()");
m.impl("topk_softmax", torch::kCUDA, &topk_softmax);
/*
* From csrc/speculative
*/
m.def(
"verify_tree_greedy(Tensor! predicts, Tensor! accept_index, Tensor! accept_token_num, "
"Tensor candidates, Tensor retrive_index, Tensor retrive_next_token, Tensor retrive_next_sibling, "

4
sgl-kernel/tests/test_per_tensor_quant_fp8.py
View File

@@ -7,8 +7,8 @@ from sgl_kernel import sgl_per_tensor_quant_fp8
from sglang.srt.utils import is_hip
is_hip_ = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
def sglang_scaled_fp8_quant(

4
sgl-kernel/tests/test_per_token_quant_fp8.py
View File

@@ -7,8 +7,8 @@ from sgl_kernel import sgl_per_token_quant_fp8
from sglang.srt.utils import is_hip
is_hip_ = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
def torch_per_token_quant_fp8(tensor, inv_scale):