sglang 0.4.5.post1__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/bert.py

@@ -0,0 +1,398 @@
+# SPDX-License-Identifier: Apache-2.0
+from typing import Any, Dict, Iterable, Optional, Set, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.distributed import get_tensor_model_parallel_world_size
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import AttentionType, RadixAttention
+from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+
+BertConfig = None
+
+
+class BertEmbedding(nn.Module):
+
+    def __init__(self, config: BertConfig):
+
+        super().__init__()
+        self.size = config.hidden_size
+        self.word_embeddings = VocabParallelEmbedding(
+            config.vocab_size, config.hidden_size
+        )
+        self.position_embeddings = VocabParallelEmbedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.token_type_embeddings = VocabParallelEmbedding(
+            config.type_vocab_size, config.hidden_size
+        )
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.position_ids = nn.Parameter(
+            torch.empty((1, config.max_position_embeddings)),
+        )
+
+        self.position_embedding_type = config.position_embedding_type
+        if self.position_embedding_type != "absolute":
+            raise ValueError(
+                "Only 'absolute' position_embedding_type" + " is supported"
+            )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+    ) -> torch.Tensor:
+        input_shape = input_ids.size()
+
+        # Input embeddings.
+        inputs_embeds = self.word_embeddings(input_ids)
+
+        # Position embeddings.
+        position_embeddings = self.position_embeddings(position_ids)
+
+        token_type_ids = torch.zeros(
+            input_shape, dtype=torch.long, device=inputs_embeds.device
+        )
+
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings + position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        return embeddings
+
+
+class BertEncoder(nn.Module):
+
+    def __init__(
+        self,
+        config: BertConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.layer = nn.ModuleList(
+            [
+                BertLayer(
+                    config=config,
+                    layer_id=layer_idx,
+                    quant_config=quant_config,
+                    prefix=f"{prefix}.layer.{layer_idx}",
+                )
+                for layer_idx in range(config.num_hidden_layers)
+            ]
+        )
+
+    def forward(
+        self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
+    ) -> torch.Tensor:
+        for layer in self.layer:
+            hidden_states = layer(hidden_states, forward_batch)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: BertConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.attention = BertAttention(
+            hidden_size=config.hidden_size,
+            num_attention_heads=config.num_attention_heads,
+            layer_id=layer_id,
+            layer_norm_eps=config.layer_norm_eps,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attention",
+        )
+
+        self.intermediate = BertIntermediate(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.intermediate",
+        )
+
+        self.output = BertOutput(
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            layer_norm_eps=config.layer_norm_eps,
+            quant_config=quant_config,
+            prefix=f"{prefix}.output",
+        )
+
+    def forward(self, hidden_states: torch.Tensor, forward_batch: ForwardBatch):
+        attn_output = self.attention(hidden_states, forward_batch)
+        intermediate_output = self.intermediate(attn_output)
+        output = self.output(intermediate_output, attn_output)
+        return output
+
+
+class BertAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_attention_heads: int,
+        layer_norm_eps: float,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.self_attn = BertSelfAttention(
+            hidden_size=hidden_size,
+            num_attention_heads=num_attention_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=f"{prefix}.output",
+        )
+
+        self.output = BertSelfOutput(
+            hidden_size=hidden_size,
+            layer_norm_eps=layer_norm_eps,
+            quant_config=quant_config,
+            prefix=f"{prefix}.output",
+        )
+
+    def forward(
+        self, hidden_states: torch.Tensor, forward_batch: ForwardBatch
+    ) -> torch.Tensor:
+        self_output = self.self_attn(hidden_states, forward_batch)
+        return self.output(self_output, hidden_states)
+
+
+class BertSelfAttention(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_attention_heads: int,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+
+        self.total_num_heads = num_attention_heads
+        assert self.total_num_heads % tp_size == 0
+
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = self.total_num_heads
+        self.head_dim = self.hidden_size // self.total_num_heads
+        assert self.head_dim * self.total_num_heads == self.hidden_size
+
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=self.hidden_size,
+            head_size=self.head_dim,
+            total_num_heads=self.total_num_heads,
+            total_num_kv_heads=self.total_num_kv_heads,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+
+        self.attn = RadixAttention(
+            num_heads=self.num_heads,
+            head_dim=self.head_dim,
+            scaling=self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            prefix=f"{prefix}.attn",
+            attn_type=AttentionType.ENCODER_ONLY,
+        )
+
+    def forward(
+        self, hidden_states: torch.Tensor, 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)
+        output = self.attn(q, k, v, forward_batch)
+        return output
+
+
+class BertSelfOutput(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        layer_norm_eps: float,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.dense = RowParallelLinear(
+            input_size=hidden_size,
+            output_size=hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.dense",
+        )
+        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states, _ = self.dense(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertIntermediate(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.dense = ColumnParallelLinear(
+            input_size=hidden_size,
+            output_size=intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.dense",
+        )
+        self.intermediate_act_fn = get_act_fn(hidden_act)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        layer_norm_eps: float,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.dense = RowParallelLinear(
+            input_size=intermediate_size,
+            output_size=hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.dense",
+        )
+
+        self.LayerNorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
+
+    def forward(
+        self, hidden_states: torch.Tensor, input_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states, _ = self.dense(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertModel(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        config: BertConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.embeddings = BertEmbedding(config)
+        self.encoder = BertEncoder(
+            config=config, quant_config=quant_config, prefix=f"encoder"
+        )
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+        # self.pooler = BertPooler(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
+    ) -> torch.Tensor:
+        assert get_embedding == True
+        # Your tokenized IDs
+
+        hidden_states = self.embeddings(
+            input_ids=input_ids,
+            position_ids=positions,
+        )
+
+        hidden_states = self.encoder(hidden_states, forward_batch=forward_batch)
+        return self.pooler(hidden_states, forward_batch)
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "query", "q"),
+            ("qkv_proj", "key", "k"),
+            ("qkv_proj", "value", "v"),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            name = name.replace("self", "self_attn")
+            if "pooler" in name:
+                continue
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+
+class Contriever(BertModel):
+    pass
+
+
+EntryClass = [BertModel, Contriever]

sglang/srt/models/deepseek.py

@@ -170,7 +170,7 @@ class DeepseekMoE(nn.Module):
             shared_output = self.shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
-        final_hidden_states = fused_moe(
+        final_hidden_states = fused_moe.fused_moe(
             hidden_states,
             self.w1,
             self.w2,

sglang/srt/models/deepseek_nextn.py

@@ -40,7 +40,7 @@ from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek_v2 import DeepseekV2DecoderLayer, DeepseekV3ForCausalLM
-from sglang.srt.utils import add_prefix, is_cuda, is_hip
+from sglang.srt.utils import BumpAllocator, add_prefix, is_cuda, is_hip
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
@@ -48,7 +48,7 @@ _is_cuda = is_cuda()
 if _is_cuda:
     from sgl_kernel import awq_dequantize
 else:
-    from vllm import _custom_ops as ops
+    from vllm._custom_ops import awq_dequantize
 
 
 class DeepseekModelNextN(nn.Module):
@@ -91,6 +91,14 @@ class DeepseekModelNextN(nn.Module):
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
+        zero_allocator = BumpAllocator(
+            buffer_size=2,
+            dtype=torch.float32,
+            device=(
+                input_embeds.device if input_embeds is not None else input_ids.device
+            ),
+        )
+
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
         else:
@@ -108,7 +116,7 @@ class DeepseekModelNextN(nn.Module):
 
         residual = None
         hidden_states, residual = self.decoder(
-            positions, hidden_states, forward_batch, residual
+            positions, hidden_states, forward_batch, residual, zero_allocator
         )
 
         if not forward_batch.forward_mode.is_idle():
@@ -262,79 +270,75 @@ class DeepseekV3ForCausalLMNextN(DeepseekV3ForCausalLM):
                     )
                     weight_loader(param, loaded_weight)
 
-        if not global_server_args_dict["disable_mla"]:
-            self_attn = self.model.decoder.self_attn
-            if hasattr(self_attn.kv_b_proj, "qweight"):
-                # AWQ compatible
-                if _is_cuda:
-                    w = awq_dequantize(
-                        self_attn.kv_b_proj.qweight,
-                        self_attn.kv_b_proj.scales,
-                        self_attn.kv_b_proj.qzeros,
-                    ).T
-                else:
-                    w = ops.awq_dequantize(
-                        self_attn.kv_b_proj.qweight,
-                        self_attn.kv_b_proj.scales,
-                        self_attn.kv_b_proj.qzeros,
-                        0,
-                        0,
-                        0,
-                    ).T
+        self_attn = self.model.decoder.self_attn
+        if hasattr(self_attn.kv_b_proj, "qweight"):
+            # AWQ compatible
+            if _is_cuda:
+                w = awq_dequantize(
+                    self_attn.kv_b_proj.qweight,
+                    self_attn.kv_b_proj.scales,
+                    self_attn.kv_b_proj.qzeros,
+                ).T
             else:
-                w = self_attn.kv_b_proj.weight
-            # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
-            # This may affect the accuracy of fp8 model.
-            if hasattr(self.quant_config, "weight_block_size") and w.dtype in (
-                torch.float8_e4m3fn,
-                torch.float8_e4m3fnuz,
-            ):
+                w = awq_dequantize(
+                    self_attn.kv_b_proj.qweight,
+                    self_attn.kv_b_proj.scales,
+                    self_attn.kv_b_proj.qzeros,
+                    0,
+                    0,
+                    0,
+                ).T
+        else:
+            w = self_attn.kv_b_proj.weight
+        # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
+        # This may affect the accuracy of fp8 model.
+        if hasattr(self.quant_config, "weight_block_size") and w.dtype in (
+            torch.float8_e4m3fn,
+            torch.float8_e4m3fnuz,
+        ):
+            weight_block_size = self.quant_config.weight_block_size
+            if weight_block_size is not None:
+                assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
+                if _is_hip:
+                    weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
+                        weight=w,
+                        weight_scale=self_attn.kv_b_proj.weight_scale_inv,
+                        input_scale=None,
+                    )
+                else:
+                    weight = w
+                    weight_scale = self_attn.kv_b_proj.weight_scale_inv
+
+                w, scale = block_quant_to_tensor_quant(
+                    weight, weight_scale, weight_block_size
+                )
+                self_attn.w_scale = scale
+        if w.dtype == torch.int8:
+            if hasattr(self.quant_config, "weight_block_size"):
+                # block-wise int8 need it
                 weight_block_size = self.quant_config.weight_block_size
                 if weight_block_size is not None:
                     assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
-                    if _is_hip:
-                        weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
-                            weight=w,
-                            weight_scale=self_attn.kv_b_proj.weight_scale_inv,
-                            input_scale=None,
-                        )
-                    else:
-                        weight = w
-                        weight_scale = self_attn.kv_b_proj.weight_scale_inv
-
-                    w, scale = block_quant_to_tensor_quant(
-                        weight, weight_scale, weight_block_size
-                    )
-                    self_attn.w_scale = scale
-            if w.dtype == torch.int8:
-                if hasattr(self.quant_config, "weight_block_size"):
-                    # block-wise int8 need it
-                    weight_block_size = self.quant_config.weight_block_size
-                    if weight_block_size is not None:
-                        assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
-                        weight = w
-                        weight_scale = self_attn.kv_b_proj.weight_scale_inv
-                        w = int8_block_dequant(
-                            weight, weight_scale, weight_block_size
-                        ).to(torch.bfloat16)
-                else:
-                    # channel-wise int8 need it
-                    assert hasattr(self_attn.kv_b_proj, "weight_scale")
-                    w = w.to(torch.bfloat16) * self_attn.kv_b_proj.weight_scale.to(
+                    weight = w
+                    weight_scale = self_attn.kv_b_proj.weight_scale_inv
+                    w = int8_block_dequant(weight, weight_scale, weight_block_size).to(
                         torch.bfloat16
                     )
-            w_kc, w_vc = w.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")
-                and self_attn.w_scale is None
-            ):
-                self_attn.w_scale = self_attn.kv_b_proj.weight_scale
-                if _is_hip:
-                    self_attn.w_scale *= 2.0
+            else:
+                # channel-wise int8 need it
+                assert hasattr(self_attn.kv_b_proj, "weight_scale")
+                w = w.to(torch.bfloat16) * self_attn.kv_b_proj.weight_scale.to(
+                    torch.bfloat16
+                )
+        w_kc, w_vc = w.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") and self_attn.w_scale is None:
+            self_attn.w_scale = self_attn.kv_b_proj.weight_scale
+            if _is_hip:
+                self_attn.w_scale *= 2.0
 
 
 EntryClass = [DeepseekV3ForCausalLMNextN]