Convert formatting to use ruff instead of yapf + isort (#26247)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 07:06:22 -07:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115205 additions and 94107 deletions
+20 -16
View File
@@ -13,8 +13,8 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
@dataclass
class MLAModules:
"""Modules used in MLA.
"""
"""Modules used in MLA."""
kv_a_layernorm: torch.nn.Module
kv_b_proj: torch.nn.Module
rotary_emb: torch.nn.Module
@@ -36,7 +36,7 @@ class MultiHeadLatentAttention(CustomOp):
because there is only one in-tree implementation in forward_native.
TODO: implement this with a new PluggableLayer mechanism.
This class takes positions and hidden_states as input.
This class takes positions and hidden_states as input.
The input tensors can either contain prefill tokens or decode tokens.
The class does the following:
@@ -125,12 +125,15 @@ class MultiHeadLatentAttention(CustomOp):
kv_lora = None
if self.q_lora_rank is not None:
assert self.fused_qkv_a_proj is not None, \
assert self.fused_qkv_a_proj is not None, (
"fused_qkv_a_proj is required when q_lora_rank is not None"
assert self.q_a_layernorm is not None, \
)
assert self.q_a_layernorm is not None, (
"q_a_layernorm is required when q_lora_rank is not None"
assert self.q_b_proj is not None, \
)
assert self.q_b_proj is not None, (
"q_b_proj is required when q_lora_rank is not None"
)
qkv_lora = self.fused_qkv_a_proj(hidden_states)[0]
q_c, kv_lora = qkv_lora.split(
[self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
@@ -139,34 +142,35 @@ class MultiHeadLatentAttention(CustomOp):
q_c = self.q_a_layernorm(q_c)
q = self.q_b_proj(q_c)[0]
else:
assert self.kv_a_proj_with_mqa is not None, \
assert self.kv_a_proj_with_mqa is not None, (
"kv_a_proj_with_mqa is required when q_lora_rank is None"
assert self.q_proj is not None, \
)
assert self.q_proj is not None, (
"q_proj is required when q_lora_rank is None"
)
kv_lora = self.kv_a_proj_with_mqa(hidden_states)[0]
q = self.q_proj(hidden_states)[0]
kv_c, k_pe = kv_lora.split([self.kv_lora_rank, self.qk_rope_head_dim],
dim=-1)
kv_c, k_pe = kv_lora.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
kv_c_normed = self.kv_a_layernorm(kv_c)
q = q.view(-1, self.num_heads, self.qk_head_dim)
# Add head dim of 1 to k_pe
k_pe = k_pe.unsqueeze(1)
q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
positions, q[..., self.qk_nope_head_dim:], k_pe)
q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb(
positions, q[..., self.qk_nope_head_dim :], k_pe
)
if self.indexer and self.is_sparse:
_topk_indices = self.indexer(hidden_states, q_c, positions,
self.rotary_emb)
_topk_indices = self.indexer(hidden_states, q_c, positions, self.rotary_emb)
attn_out = self.mla_attn(
q,
kv_c_normed,
k_pe,
output_shape=(hidden_states.shape[0],
self.num_heads * self.v_head_dim))
output_shape=(hidden_states.shape[0], self.num_heads * self.v_head_dim),
)
return self.o_proj(attn_out)[0]
def forward_cuda(self, *args, **kwargs):