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

[Model][2/N] Remove deepseek_mtp modeling. (#3561)

Browse Source 
This PR is step 2 of deepseek model refactoring and removes
deepseek_mtp.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-10-21 20:17:09 +08:00
committed by GitHub
parent ffb42a8daa
commit 220df60c61
7 changed files with 38 additions and 422 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
tests/ut/attention/test_mla_v1.py
View File

@@ -306,6 +306,7 @@ class TestAscendMLAImpl(TestBase):
"kv_b_proj": MagicMock(),
"o_proj": MagicMock(),
"kv_a_proj_with_mqa": MagicMock(),
"fused_qkv_a_proj": MagicMock(),
"kv_a_layernorm": kv_a_layernorm,
}
@@ -511,7 +512,6 @@ class TestAscendMLAImpl(TestBase):
attn_metadata.prefill.cos = torch.randn(2, 64)
attn_metadata.prefill.sin = torch.randn(2, 64)
self.impl.q_a_proj = MagicMock()
self.impl.q_a_layernorm = MagicMock()
self.impl.q_a_layernorm.return_value = torch.randn(
attn_metadata.num_actual_tokens, self.impl.num_heads,
@@ -519,7 +519,14 @@ class TestAscendMLAImpl(TestBase):
self.impl.kv_a_proj_with_mqa = MagicMock()
self.impl.kv_a_proj_with_mqa.return_value = [
torch.randn(num_prefill_tokens, self.impl.num_heads,
self.impl.qk_nope_head_dim + self.impl.kv_lora_rank)
self.impl.qk_rope_head_dim + self.impl.kv_lora_rank)
]
self.impl.fused_qkv_a_proj = MagicMock()
self.impl.fused_qkv_a_proj.return_value = [
torch.randn(
num_prefill_tokens, self.impl.num_heads,
self.impl.qk_rope_head_dim + self.impl.kv_lora_rank +
self.impl.q_lora_rank)
]
self.impl.q_proj = MagicMock()
self.impl.q_proj.return_value = [