[3/n] deepseek_v2.py Refactor: Migrate MLA forward method in deepseek_v2.py (#19122)

This commit is contained in:
Baizhou Zhang
2026-02-27 13:37:29 -08:00
committed by GitHub
parent 7e46aafebb
commit 776709efe8
9 changed files with 906 additions and 818 deletions
+3 -3
View File
@@ -8,7 +8,7 @@ from utils import (
precision,
)
from sglang.srt.layers.rotary_embedding import _apply_rotary_emb
from sglang.srt.layers.rotary_embedding.utils import apply_rotary_emb
from sglang.test.test_utils import CustomTestCase
convert_weight_packed = torch.ops.sgl_kernel.convert_weight_packed
@@ -46,8 +46,8 @@ def rotary_emb(q_pe, k_pe, pos, cos_sin_cache):
key_rot = k_pe[..., :rotary_dim]
cos_sin = cos_sin_cache[pos]
cos, sin = cos_sin.chunk(2, dim=-1)
query_rot = _apply_rotary_emb(query_rot, cos, sin, False)
key_rot = _apply_rotary_emb(key_rot, cos, sin, False)
query_rot = apply_rotary_emb(query_rot, cos, sin, False)
key_rot = apply_rotary_emb(key_rot, cos, sin, False)
return query_rot.to(orig_dtype), key_rot.to(orig_dtype)
+2 -2
View File
@@ -3,9 +3,9 @@ import unittest
import torch
from utils import precision
from sglang.srt.layers.rotary_embedding import (
from sglang.srt.layers.rotary_embedding.base import RotaryEmbedding
from sglang.srt.layers.rotary_embedding.rope_variant import (
DeepseekScalingRotaryEmbedding,
RotaryEmbedding,
)
from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
from sglang.test.test_utils import CustomTestCase