Skip to content

[docker] Fix IndexCache with mla model #1736

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zhuzilin merged 1 commit into from
Mar 18, 2026
Merged

[docker] Fix IndexCache with mla model #1736

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 27 additions & 21 deletions docker/patch/latest/sglang.patch
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2518,7 +2518,7 @@ index cb13a7c67..29cd51d4a 100644
hidden_states,
forward_batch,
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 1583dd788..432758732 100644
index 1583dd788..08d59c883 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -1085,6 +1085,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
Expand All @@ -2529,14 +2529,20 @@ index 1583dd788..432758732 100644
) -> None:
super().__init__()
self.layer_id = layer_id
@@ -1174,6 +1175,34 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1154,6 +1155,8 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
prefix=add_prefix("kv_a_proj_with_mqa", prefix),
)

+ self.skip_topk = False
+ self.next_skip_topk = False
if self.use_nsa:
is_neox_style = not getattr(config, "indexer_rope_interleave", False)
self.indexer = Indexer(
@@ -1174,6 +1177,31 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
layer_id=layer_id,
alt_stream=alt_stream,
)
+ if is_nextn:
+ self.skip_topk = False
+ self.next_skip_topk = False
+ else:
+ if not is_nextn:
+ self.index_topk_freq = getattr(config, "index_topk_freq", 1)
+ self.index_topk_pattern = getattr(config, "index_topk_pattern", None)
+ self.index_skip_topk_offset = getattr(
Expand Down Expand Up @@ -2564,31 +2570,31 @@ index 1583dd788..432758732 100644

self.kv_b_proj = ColumnParallelLinear(
self.kv_lora_rank,
@@ -1362,6 +1391,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1362,6 +1390,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
forward_batch: ForwardBatch,
zero_allocator: BumpAllocator,
llama_4_scaling: Optional[torch.Tensor] = None,
+ prev_topk_indices: Optional[torch.Tensor] = None,
):
s = self.forward_prepare(
positions=positions,
@@ -1369,6 +1399,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1369,6 +1398,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
forward_batch=forward_batch,
zero_allocator=zero_allocator,
llama_4_scaling=llama_4_scaling,
+ prev_topk_indices=prev_topk_indices,
)
return self.forward_core(s)

@@ -1379,6 +1410,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1379,6 +1409,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
forward_batch: ForwardBatch,
zero_allocator: BumpAllocator,
llama_4_scaling: Optional[torch.Tensor] = None,
+ prev_topk_indices: Optional[torch.Tensor] = None,
):
if self.attn_mha.kv_b_proj is None:
self.attn_mha.kv_b_proj = self.kv_b_proj
@@ -1418,7 +1450,12 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1418,7 +1449,12 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
)
elif attn_forward_method == AttnForwardMethod.MLA:
inner_state = self.forward_absorb_prepare(
Expand All @@ -2602,15 +2608,15 @@ index 1583dd788..432758732 100644
)
elif attn_forward_method == AttnForwardMethod.MLA_FUSED_ROPE:
inner_state = self.forward_absorb_fused_mla_rope_prepare(
@@ -1529,6 +1566,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1529,6 +1565,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
forward_batch: ForwardBatch,
zero_allocator: BumpAllocator,
llama_4_scaling: Optional[torch.Tensor] = None,
+ prev_topk_indices: Optional[torch.Tensor] = None,
):
from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode

@@ -1620,18 +1658,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1620,18 +1657,7 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
q = self.q_b_proj(q)[0].view(
-1, self.num_local_heads, self.qk_head_dim
)
Expand All @@ -2630,7 +2636,7 @@ index 1583dd788..432758732 100644
topk_indices = self.indexer(
x=hidden_states,
q_lora=q_lora,
@@ -1639,6 +1666,23 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1639,6 +1665,23 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
forward_batch=forward_batch,
layer_id=self.layer_id,
)
Expand All @@ -2654,7 +2660,7 @@ index 1583dd788..432758732 100644
else:
q = self.q_proj(hidden_states)[0].view(
-1, self.num_local_heads, self.qk_head_dim
@@ -1929,8 +1973,10 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
@@ -1929,8 +1972,10 @@ class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
).transpose(0, 1),
)
output, _ = self.o_proj(attn_bmm_output)
Expand All @@ -2667,23 +2673,23 @@ index 1583dd788..432758732 100644

def forward_absorb_fused_mla_rope_prepare(
self,
@@ -2275,6 +2321,7 @@ class DeepseekV2DecoderLayer(nn.Module):
@@ -2275,6 +2320,7 @@ class DeepseekV2DecoderLayer(nn.Module):
reduce_results=False,
prefix=add_prefix("self_attn", prefix),
alt_stream=alt_stream,
+ is_nextn=is_nextn,
)

self.is_layer_sparse = self._is_layer_sparse(layer_id, is_nextn=is_nextn)
@@ -2357,6 +2404,7 @@ class DeepseekV2DecoderLayer(nn.Module):
@@ -2357,6 +2403,7 @@ class DeepseekV2DecoderLayer(nn.Module):
zero_allocator: BumpAllocator,
gemm_output_zero_allocator: BumpAllocator = None,
llama_4_scaling: Optional[torch.Tensor] = None,
+ prev_topk_indices: Optional[torch.Tensor] = None,
) -> torch.Tensor:
quant_format = (
"mxfp4"
@@ -2398,7 +2446,12 @@ class DeepseekV2DecoderLayer(nn.Module):
@@ -2398,7 +2445,12 @@ class DeepseekV2DecoderLayer(nn.Module):
forward_batch=forward_batch,
zero_allocator=zero_allocator,
llama_4_scaling=llama_4_scaling,
Expand All @@ -2696,7 +2702,7 @@ index 1583dd788..432758732 100644

hidden_states, residual = self.layer_communicator.prepare_mlp(
hidden_states, residual, forward_batch
@@ -2434,7 +2487,7 @@ class DeepseekV2DecoderLayer(nn.Module):
@@ -2434,7 +2486,7 @@ class DeepseekV2DecoderLayer(nn.Module):
hidden_states, residual, forward_batch
)

Expand All @@ -2705,15 +2711,15 @@ index 1583dd788..432758732 100644

def op_comm_prepare_attn(
self,
@@ -2710,6 +2763,7 @@ class DeepseekV2Model(nn.Module):
@@ -2710,6 +2762,7 @@ class DeepseekV2Model(nn.Module):
elif self.first_k_dense_replace < normal_start_layer:
normal_end_layer = normal_start_layer = 0
aux_hidden_states = []
+ topk_indices = None
for i in range(normal_start_layer, normal_end_layer):
# NOTE: torch dynamo does not support graph break in context manager
ctx = (
@@ -2727,7 +2781,7 @@ class DeepseekV2Model(nn.Module):
@@ -2727,7 +2780,7 @@ class DeepseekV2Model(nn.Module):
else:
aux_hidden_states.append(hidden_states + residual)
layer = self.layers[i]
Expand All @@ -2722,7 +2728,7 @@ index 1583dd788..432758732 100644
positions,
hidden_states,
forward_batch,
@@ -2735,6 +2789,7 @@ class DeepseekV2Model(nn.Module):
@@ -2735,6 +2788,7 @@ class DeepseekV2Model(nn.Module):
zero_allocator,
gemm_output_zero_allocator,
llama_4_scaling,
Expand Down
2 changes: 1 addition & 1 deletion docker/version.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1 +1 @@
nightly-dev-20260318a
nightly-dev-20260318b