sglang 0.4.5__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/exaone.py

@@ -155,6 +155,7 @@ class ExaoneAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
         )
 
     def forward(

sglang/srt/models/gemma.py

@@ -137,6 +137,7 @@ class GemmaAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/gemma2.py

@@ -163,6 +163,7 @@ class Gemma2Attention(nn.Module):
                 if use_sliding_window
                 else None
             ),
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/gemma3_causal.py

@@ -193,6 +193,7 @@ class Gemma3Attention(nn.Module):
             # Module must also define `get_attention_sliding_window_size` to correctly initialize
             # attention backend in `ForwardBatch`.
             sliding_window_size=self.sliding_window,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/gpt2.py

@@ -78,6 +78,7 @@ class GPT2Attention(nn.Module):
             scaling=self.scale,
             num_kv_heads=total_num_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
         )
 
     def forward(

sglang/srt/models/gpt_bigcode.py

@@ -87,6 +87,7 @@ class GPTBigCodeAttention(nn.Module):
             scaling=self.scale,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/granite.py

@@ -158,6 +158,7 @@ class GraniteAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/grok.py

@@ -215,6 +215,7 @@ class Grok1Attention(nn.Module):
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
             logit_cap=logit_cap,
+            quant_config=quant_config,
         )
 
     def forward(

sglang/srt/models/internlm2.py

@@ -145,6 +145,7 @@ class InternLM2Attention(nn.Module):
             self.scaling,
             self.num_kv_heads,
             layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/llama.py

@@ -170,6 +170,7 @@ class LlamaAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 
@@ -361,11 +362,11 @@ class LlamaForCausalLM(nn.Module):
     column_parallel_weights_modules = [".down_proj.", ".o_proj."]
     bitsandbytes_stacked_params_mapping = {
         # shard_name, weight_name, index
-        "q_proj": ("qkv_proj", 0),
-        "k_proj": ("qkv_proj", 1),
-        "v_proj": ("qkv_proj", 2),
-        "gate_proj": ("gate_up_proj", 0),
-        "up_proj": ("gate_up_proj", 1),
+        ".q_proj": (".qkv_proj", 0),
+        ".k_proj": (".qkv_proj", 1),
+        ".v_proj": (".qkv_proj", 2),
+        ".gate_proj": (".gate_up_proj", 0),
+        ".up_proj": (".gate_up_proj", 1),
     }
 
     def __init__(

sglang/srt/models/llama4.py

@@ -27,6 +27,13 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
+from sglang.srt.layers.dp_attention import (
+    dp_gather_partial,
+    dp_scatter,
+    get_attention_dp_size,
+    get_attention_tp_rank,
+    get_attention_tp_size,
+)
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
     QKVParallelLinear,
@@ -38,9 +45,10 @@ from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.llama import LlamaForCausalLM, LlamaMLP
-from sglang.srt.utils import add_prefix, get_compiler_backend, make_layers
+from sglang.srt.utils import add_prefix, fast_topk, get_compiler_backend, make_layers
 
 logger = logging.getLogger(__name__)
 
@@ -55,7 +63,7 @@ class Llama4MoE(nn.Module):
         topk: int,
         renormalize: bool,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        router_scores_aK, router_indices_aK = torch.topk(gating_output, topk, dim=-1)
+        router_scores_aK, router_indices_aK = fast_topk(gating_output, topk, dim=-1)
         router_scores_aK = torch.sigmoid(router_scores_aK.float()).to(
             hidden_states.dtype
         )
@@ -143,20 +151,24 @@ class Llama4Attention(nn.Module):
         self.hidden_size = hidden_size
         self.use_rope = int((layer_id + 1) % 4 != 0)
         self.use_qk_norm = config.use_qk_norm and self.use_rope
-        tp_size = get_tensor_model_parallel_world_size()
+
+        self.dp_size = get_attention_dp_size()
+        attn_tp_rank = get_attention_tp_rank()
+        attn_tp_size = get_attention_tp_size()
+
         self.total_num_heads = num_heads
-        assert self.total_num_heads % tp_size == 0
-        self.num_heads = self.total_num_heads // tp_size
+        assert self.total_num_heads % attn_tp_size == 0
+        self.num_heads = self.total_num_heads // attn_tp_size
         self.total_num_kv_heads = num_kv_heads
-        if self.total_num_kv_heads >= tp_size:
+        if self.total_num_kv_heads >= attn_tp_size:
             # Number of KV heads is greater than TP size, so we partition
             # the KV heads across multiple tensor parallel GPUs.
-            assert self.total_num_kv_heads % tp_size == 0
+            assert self.total_num_kv_heads % attn_tp_size == 0
         else:
             # Number of KV heads is less than TP size, so we replicate
             # the KV heads across multiple tensor parallel GPUs.
-            assert tp_size % self.total_num_kv_heads == 0
-        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+            assert attn_tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // attn_tp_size)
         self.head_dim = config.head_dim
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
@@ -183,6 +195,8 @@ class Llama4Attention(nn.Module):
             bias=bias,
             quant_config=quant_config,
             prefix=add_prefix("qkv_proj", prefix),
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
         )
 
         self.o_proj = RowParallelLinear(
@@ -191,6 +205,9 @@ class Llama4Attention(nn.Module):
             bias=bias_o_proj,
             quant_config=quant_config,
             prefix=add_prefix("o_proj", prefix),
+            tp_rank=attn_tp_rank,
+            tp_size=attn_tp_size,
+            reduce_results=False,
         )
         is_neox_style = True
         is_gguf = quant_config and quant_config.get_name() == "gguf"
@@ -223,9 +240,13 @@ class Llama4Attention(nn.Module):
     def _get_attn_scale(self, positions: torch.Tensor) -> torch.Tensor:
         floor = torch.floor((positions + 1.0) / self.floor_scale)
         attn_scale = torch.log(floor + 1.0) * self.attn_scale + 1.0
-
         return attn_scale.unsqueeze(-1)
 
+    @torch.compile(dynamic=True, backend=get_compiler_backend())
+    def _mul_attn_scale(self, positions, q):
+        attn_scale = self._get_attn_scale(positions)
+        return (q * attn_scale).to(q.dtype)
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -233,27 +254,29 @@ class Llama4Attention(nn.Module):
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
-        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+
+        qk, v = qkv.split([self.q_size + self.kv_size, self.kv_size], dim=-1)
 
         if self.rotary_emb is not None:
-            q, k = self.rotary_emb(positions, q, k)
+            q_view, k_view = qk.split([self.q_size, self.kv_size], dim=-1)
+            q_out_unused, k_out_unused = self.rotary_emb(positions, q_view, k_view)
+            assert (q_out_unused is q_view) and (k_out_unused is k_view)
+            del q_view, k_view, q_out_unused, k_out_unused
 
         if self.qk_norm is not None:
-            # TODO: support float
-            q = q.reshape(-1, self.head_dim).contiguous().bfloat16()
-            k = k.reshape(-1, self.head_dim).contiguous().bfloat16()
-            q = self.qk_norm(q).to(q.dtype)
-            k = self.qk_norm(k).to(k.dtype)
-            q = q.reshape(-1, self.q_size)
-            k = k.reshape(-1, self.kv_size)
+            # TODO there are still 2 redundant direct_copy_kernel_cuda for this `reshape` and (in attn backend) q.contiguous(), maybe we can fuse them later
+            qk = qk.reshape(-1, self.head_dim).contiguous().bfloat16()
+            qk = self.qk_norm(qk).to(torch.bfloat16)
+            qk = qk.reshape(-1, self.q_size + self.kv_size)
+
+        q, k = qk.split([self.q_size, self.kv_size], dim=-1)
 
         # We are applying temperature tuning (https://arxiv.org/abs/2501.19399) to NoPE layers, where
         # the inference-time temperature tuning function is customized to not affect short context
         # while working at very long context
         # https://arxiv.org/abs/2501.19399
         if self.attn_temperature_tuning and not self.use_rope:
-            attn_scale = self._get_attn_scale(positions)
-            q = (q * attn_scale).to(q.dtype)
+            q = self._mul_attn_scale(positions=positions, q=q)
 
         attn_output = self.attn(q, k, v, forward_batch)
         output, _ = self.o_proj(attn_output)
@@ -274,6 +297,9 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
+        self.dp_size = get_attention_dp_size()
+        self.attn_tp_size = get_attention_tp_size()
+        self.attn_tp_rank = get_attention_tp_rank()
 
         self.self_attn = Llama4Attention(
             config=config,
@@ -316,21 +342,58 @@ class Llama4DecoderLayer(nn.Module):
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        # Self Attention
-        if residual is None:
+        if hidden_states.shape[0] == 0:
             residual = hidden_states
-            hidden_states = self.input_layernorm(hidden_states)
         else:
-            hidden_states, residual = self.input_layernorm(hidden_states, residual)
-        hidden_states = self.self_attn(
-            positions=positions,
-            hidden_states=hidden_states,
-            forward_batch=forward_batch,
-        )
+            # Self Attention
+            if residual is None:
+                residual = hidden_states
+                hidden_states = self.input_layernorm(hidden_states)
+            else:
+                hidden_states, residual = self.input_layernorm(hidden_states, residual)
+            hidden_states = self.self_attn(
+                positions=positions,
+                hidden_states=hidden_states,
+                forward_batch=forward_batch,
+            )
+
+        # Gather
+        if get_tensor_model_parallel_world_size() > 1:
+            # all gather and all reduce
+            if self.dp_size != 1:
+                if self.attn_tp_rank == 0:
+                    hidden_states += residual
+                hidden_states, local_hidden_states = (
+                    forward_batch.gathered_buffer,
+                    hidden_states,
+                )
+                dp_gather_partial(hidden_states, local_hidden_states, forward_batch)
+                dp_scatter(residual, hidden_states, forward_batch)
+                hidden_states = self.post_attention_layernorm(hidden_states)
+            else:
+                hidden_states = tensor_model_parallel_all_reduce(hidden_states)
+                hidden_states, residual = self.post_attention_layernorm(
+                    hidden_states, residual
+                )
+        else:
+            hidden_states, residual = self.post_attention_layernorm(
+                hidden_states, residual
+            )
 
         # Fully Connected
-        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
         hidden_states = self.feed_forward(hidden_states)
+
+        # TODO(ch-wan): ues reduce-scatter in MLP to avoid this scatter
+        # Scatter
+        if self.dp_size != 1:
+            # important: forward batch.gathered_buffer is used both after scatter and after gather.
+            # be careful about this!
+            hidden_states, global_hidden_states = (
+                forward_batch.gathered_buffer[: forward_batch.input_ids.shape[0]],
+                hidden_states,
+            )
+            dp_scatter(hidden_states, global_hidden_states, forward_batch)
+
         return hidden_states, residual
 
 
@@ -350,13 +413,14 @@ class Llama4Model(nn.Module):
             config.hidden_size,
             quant_config=quant_config,
             prefix=add_prefix("embed_tokens", prefix),
+            enable_tp=not global_server_args_dict["enable_dp_attention"],
         )
         self.layers = make_layers(
             config.num_hidden_layers,
             lambda idx, prefix: Llama4DecoderLayer(
                 config=config, layer_id=idx, quant_config=quant_config, prefix=prefix
             ),
-            prefix="model.layers",
+            prefix=add_prefix("layers", prefix),
         )
 
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -385,7 +449,8 @@ class Llama4Model(nn.Module):
                 forward_batch,
                 residual,
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+        if not forward_batch.forward_mode.is_idle():
+            hidden_states, _ = self.norm(hidden_states, residual)
 
         if len(aux_hidden_states) == 0:
             return hidden_states
@@ -394,7 +459,6 @@ class Llama4Model(nn.Module):
 
 
 class Llama4ForCausalLM(LlamaForCausalLM):
-
     packed_modules_mapping = {
         "qkv_proj": ["q_proj", "k_proj", "v_proj"],
         "gate_up_proj": ["gate_proj", "up_proj"],
@@ -408,6 +472,9 @@ class Llama4ForCausalLM(LlamaForCausalLM):
     ):
         super().__init__(config, quant_config, prefix)
 
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
     def _init_model(
         self,
         config: Llama4TextConfig,

sglang/srt/models/minicpm.py

@@ -146,6 +146,7 @@ class MiniCPMAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/minicpm3.py

@@ -93,157 +93,6 @@ def input_to_float8(x, dtype=torch.float8_e4m3fn):
     return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
 
 
-class MiniCPM3Attention(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        hidden_size: int,
-        num_heads: int,
-        qk_nope_head_dim: int,
-        qk_rope_head_dim: int,
-        v_head_dim: int,
-        q_lora_rank: int,
-        kv_lora_rank: int,
-        rope_theta: float = 10000,
-        rope_scaling: Optional[Dict[str, Any]] = None,
-        max_position_embeddings: int = 8192,
-        quant_config: Optional[QuantizationConfig] = None,
-        layer_id=None,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        self.layer_id = layer_id
-        self.hidden_size = hidden_size
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.qk_rope_head_dim = qk_rope_head_dim
-        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
-        self.v_head_dim = v_head_dim
-        self.q_lora_rank = q_lora_rank
-        self.kv_lora_rank = kv_lora_rank
-        self.num_heads = num_heads
-        tp_size = get_tensor_model_parallel_world_size()
-        assert num_heads % tp_size == 0
-        self.num_local_heads = num_heads // tp_size
-        self.scaling = self.qk_head_dim**-0.5
-        self.rope_theta = rope_theta
-        self.max_position_embeddings = max_position_embeddings
-
-        if self.q_lora_rank is not None:
-            self.q_a_proj = ReplicatedLinear(
-                self.hidden_size,
-                self.q_lora_rank,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_a_proj", prefix),
-            )
-            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
-            self.q_b_proj = ColumnParallelLinear(
-                q_lora_rank,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_b_proj", prefix),
-            )
-        else:
-            self.q_proj = ColumnParallelLinear(
-                self.hidden_size,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_proj", prefix),
-            )
-
-        self.kv_a_proj_with_mqa = ReplicatedLinear(
-            self.hidden_size,
-            self.kv_lora_rank + self.qk_rope_head_dim,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_a_proj_with_mqa", prefix),
-        )
-        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
-        self.kv_b_proj = ColumnParallelLinear(
-            self.kv_lora_rank,
-            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_b_proj", prefix),
-        )
-        # O projection.
-        self.o_proj = RowParallelLinear(
-            self.num_heads * self.v_head_dim,
-            self.hidden_size,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("o_proj", prefix),
-        )
-        self.rotary_emb = get_rope(
-            qk_rope_head_dim,
-            rotary_dim=qk_rope_head_dim,
-            max_position=max_position_embeddings,
-            base=rope_theta,
-            rope_scaling=rope_scaling,
-        )
-
-        # TODO support head_size 96
-        self.attn = RadixAttention(
-            self.num_local_heads,
-            128,
-            self.scaling,
-            num_kv_heads=self.num_local_heads,
-            layer_id=layer_id,
-            prefix=add_prefix("attn", prefix),
-        )
-
-    def forward(
-        self,
-        positions: torch.Tensor,
-        hidden_states: torch.Tensor,
-        forward_batch: ForwardBatch,
-    ) -> torch.Tensor:
-        if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
-            q = self.q_a_layernorm(q)
-            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
-        else:
-            q = self.q_proj(hidden_states)[0].view(
-                -1, self.num_local_heads, self.qk_head_dim
-            )
-        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
-        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
-        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
-        latent_cache = latent_cache.unsqueeze(1)
-        kv_a = self.kv_a_layernorm(kv_a.contiguous())
-        kv = self.kv_b_proj(kv_a)[0]
-        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
-        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
-        k_pe = latent_cache[:, :, self.kv_lora_rank :]
-        original_shapes = [q_pe.shape, k_pe.shape]
-        q_pe, k_pe = self.rotary_emb(
-            positions, q_pe.reshape(q_pe.shape[0], -1), k_pe.reshape(k_pe.shape[0], -1)
-        )
-        q_pe, k_pe = q_pe.view(original_shapes[0]), k_pe.view(original_shapes[1])
-        q[..., self.qk_nope_head_dim :] = q_pe
-        k = torch.empty_like(q)
-        k[..., : self.qk_nope_head_dim] = k_nope
-        k[..., self.qk_nope_head_dim :] = k_pe
-        q = torch.nn.functional.pad(q, [0, 128 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        k = torch.nn.functional.pad(k, [0, 128 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        v = torch.nn.functional.pad(v, [0, 128 - self.v_head_dim], value=0).view(
-            -1, self.num_local_heads * 128
-        )
-        attn_output = self.attn(q, k, v, forward_batch)
-        attn_output = attn_output.view(-1, self.num_local_heads, 128)[
-            ..., : self.v_head_dim
-        ].reshape(-1, self.num_local_heads * self.v_head_dim)
-        output, _ = self.o_proj(attn_output)
-        return output
-
-
 class MiniCPM3AttentionMLA(nn.Module):
 
     def __init__(
@@ -343,6 +192,7 @@ class MiniCPM3AttentionMLA(nn.Module):
             num_kv_heads=1,
             layer_id=layer_id,
             v_head_dim=self.kv_lora_rank,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 
@@ -432,44 +282,25 @@ class MiniCPM3DecoderLayer(nn.Module):
         rope_theta = getattr(config, "rope_theta", 10000)
         rope_scaling = getattr(config, "rope_scaling", None)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
-        if not global_server_args_dict["disable_mla"]:
-            self.self_attn = MiniCPM3AttentionMLA(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=self.hidden_size // config.num_attention_heads,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                prefix=add_prefix("self_attn", prefix),
-            )
-        else:
-            self.self_attn = MiniCPM3Attention(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=self.hidden_size // config.num_attention_heads,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                prefix=add_prefix("self_attn", prefix),
-            )
+        self.self_attn = MiniCPM3AttentionMLA(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            qk_nope_head_dim=config.qk_nope_head_dim,
+            qk_rope_head_dim=config.qk_rope_head_dim,
+            v_head_dim=self.hidden_size // config.num_attention_heads,
+            q_lora_rank=(
+                config.q_lora_rank if hasattr(config, "q_lora_rank") else None
+            ),
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            layer_id=layer_id,
+            prefix=add_prefix("self_attn", prefix),
+        )
+
         self.mlp = MiniCPM3MLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
@@ -672,17 +503,16 @@ class MiniCPM3ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
 
-        if not global_server_args_dict["disable_mla"]:
-            for layer_id in range(self.config.num_hidden_layers):
-                self_attn = self.model.layers[layer_id].self_attn
-                w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(
-                    0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
-                ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
-                self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
-                self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
-                if hasattr(self_attn.kv_b_proj, "weight_scale"):
-                    self_attn.w_scale = self_attn.kv_b_proj.weight_scale
-                del self_attn.kv_b_proj
+        for layer_id in range(self.config.num_hidden_layers):
+            self_attn = self.model.layers[layer_id].self_attn
+            w_kc, w_vc = self_attn.kv_b_proj.weight.unflatten(
+                0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
+            ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
+            self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
+            self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
+            if hasattr(self_attn.kv_b_proj, "weight_scale"):
+                self_attn.w_scale = self_attn.kv_b_proj.weight_scale
+            del self_attn.kv_b_proj
 
 
 EntryClass = MiniCPM3ForCausalLM

sglang/srt/models/mixtral.py

@@ -169,6 +169,7 @@ class MixtralAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/mixtral_quant.py

@@ -232,6 +232,7 @@ class MixtralAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )