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

Clean up imports (#5467)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-04-16 15:26:49 -07:00
committed by GitHub
parent d7bc19a46a
commit 177320a582
51 changed files with 376 additions and 573 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/models/deepseek_nextn.py
View File

@@ -48,7 +48,7 @@ _is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import awq_dequantize
else:
from vllm import _custom_ops as ops
from vllm._custom_ops import awq_dequantize
class DeepseekModelNextN(nn.Module):
@@ -273,7 +273,7 @@ class DeepseekV3ForCausalLMNextN(DeepseekV3ForCausalLM):
self_attn.kv_b_proj.qzeros,
).T
else:
w = ops.awq_dequantize(
w = awq_dequantize(
self_attn.kv_b_proj.qweight,
self_attn.kv_b_proj.scales,
self_attn.kv_b_proj.qzeros,

11
python/sglang/srt/models/deepseek_v2.py
View File

@@ -51,6 +51,7 @@ from sglang.srt.layers.linear import (
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, EPMoE
from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -80,10 +81,8 @@ _is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import awq_dequantize, bmm_fp8, merge_state_v2
from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
else:
from vllm import _custom_ops as ops
from vllm._custom_ops import awq_dequantize
if _is_hip:
from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
@@ -861,7 +860,7 @@ class DeepseekV2AttentionMLA(nn.Module):
)
elif self.w_kc.dtype == torch.float8_e4m3fn:
q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
q_nope.transpose(0, 1), dtype=torch.float8_e4m3fn
q_nope.transpose(0, 1),
)
q_nope_out = bmm_fp8(
q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
@@ -892,7 +891,7 @@ class DeepseekV2AttentionMLA(nn.Module):
)
elif self.w_vc.dtype == torch.float8_e4m3fn:
attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
attn_output.transpose(0, 1), dtype=torch.float8_e4m3fn
attn_output.transpose(0, 1),
)
attn_bmm_output = bmm_fp8(
attn_output_val,
@@ -1565,7 +1564,7 @@ class DeepseekV2ForCausalLM(nn.Module):
self_attn.kv_b_proj.qzeros,
).T
else:
w = ops.awq_dequantize(
w = awq_dequantize(
self_attn.kv_b_proj.qweight,
self_attn.kv_b_proj.scales,
self_attn.kv_b_proj.qzeros,