sglang 0.5.4__py3-none-any.whl → 0.5.4.post2__py3-none-any.whl

sglang/srt/models/glm4.py

@@ -15,46 +15,119 @@
 # Modeling from:
 # ./llama.py and
 # https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4/modular_glm4.py
-"""Inference-only GLM4 model compatible with THUDM weights."""
+"""Inference-only GLM-4-0414 model compatible with HuggingFace weights."""
 
-from typing import Iterable, List, Optional, Tuple, Union
+import logging
+from typing import Any, Dict, Iterable, Optional, Tuple, Union
 
 import torch
 from torch import nn
-from transformers import Glm4Config
 
-from sglang.srt.distributed import get_tensor_model_parallel_world_size
+from sglang.srt.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.dp_attention import is_dp_attention_enabled
 from sglang.srt.layers.layernorm import RMSNorm
-from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.pooler import Pooler, PoolingType
 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.utils import PPMissingLayer, get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-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.llama import LlamaMLP as Glm4MLP
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    kv_cache_scales_loader,
+)
 from sglang.srt.utils import add_prefix, make_layers
 
+Glm4Config = None
+
+logger = logging.getLogger(__name__)
+
+
+class Glm4MLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        reduce_results: bool = True,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("gate_up_proj", prefix),
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
+            reduce_results=reduce_results,
+        )
+        if hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported activation: {hidden_act}. Only silu is supported for now."
+            )
+        self.act_fn = SiluAndMul()
+
+    def forward(
+        self,
+        x,
+        forward_batch=None,
+        use_reduce_scatter: bool = False,
+    ):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(
+            x,
+            skip_all_reduce=use_reduce_scatter,
+        )
+        return x
+
 
 class Glm4Attention(nn.Module):
     def __init__(
         self,
-        config,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        head_dim: Optional[int] = None,
         layer_id: int = 0,
+        rope_theta: float = 1000000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 131072,
         quant_config: Optional[QuantizationConfig] = None,
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
+        partial_rotary_factor: float = 0.5,
         prefix: str = "",
-    ):
+    ) -> None:
         super().__init__()
-        self.hidden_size = config.hidden_size
+        self.hidden_size = hidden_size
         tp_size = get_tensor_model_parallel_world_size()
-        self.total_num_heads = config.num_attention_heads
+        self.total_num_heads = num_heads
         assert self.total_num_heads % tp_size == 0
         self.num_heads = self.total_num_heads // tp_size
-        self.total_num_kv_heads = config.num_key_value_heads
+        self.total_num_kv_heads = num_kv_heads
         if self.total_num_kv_heads >= tp_size:
             # Number of KV heads is greater than TP size, so we partition
             # the KV heads across multiple tensor parallel GPUs.
@@ -63,27 +136,30 @@ class Glm4Attention(nn.Module):
             # 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
-        partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
         self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
-        self.head_dim = config.hidden_size // self.total_num_heads
+        if head_dim is not None:
+            self.head_dim = head_dim
+        else:
+            self.head_dim = hidden_size // self.total_num_heads
         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.rope_theta = getattr(config, "rope_theta", 1000000)
-        self.rope_scaling = getattr(config, "rope_scaling", None)
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+        self.partial_rotary_factor = partial_rotary_factor
 
         self.qkv_proj = QKVParallelLinear(
-            self.hidden_size,
+            hidden_size,
             self.head_dim,
             self.total_num_heads,
             self.total_num_kv_heads,
-            bias=config.attention_bias,
+            bias=True,
             quant_config=quant_config,
             prefix=add_prefix("qkv_proj", prefix),
         )
         self.o_proj = RowParallelLinear(
             self.total_num_heads * self.head_dim,
-            self.hidden_size,
+            hidden_size,
             bias=False,
             quant_config=quant_config,
             prefix=add_prefix("o_proj", prefix),
@@ -92,9 +168,10 @@ class Glm4Attention(nn.Module):
         self.rotary_emb = get_rope(
             self.head_dim,
             rotary_dim=self.head_dim,
-            max_position=config.max_position_embeddings,
-            base=self.rope_theta,
-            rope_scaling=self.rope_scaling,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            dual_chunk_attention_config=dual_chunk_attention_config,
             partial_rotary_factor=partial_rotary_factor,
             is_neox_style=False,
         )
@@ -117,14 +194,9 @@ class Glm4Attention(nn.Module):
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
         q, k = self.rotary_emb(positions, q, k)
-        context_layer = self.attn(
-            q,
-            k,
-            v,
-            forward_batch,
-        )
-        attn_output, _ = self.o_proj(context_layer)
-        return attn_output
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
 
 
 class Glm4DecoderLayer(nn.Module):
@@ -136,15 +208,35 @@ class Glm4DecoderLayer(nn.Module):
 
     def __init__(
         self,
-        config,
-        layer_id: int,
+        config: Glm4Config,
+        layer_id: int = 0,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
-    ):
+        alt_stream: Optional[torch.cuda.Stream] = None,
+    ) -> None:
         super().__init__()
-        # Self attention.
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 1000000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
+        head_dim = getattr(config, "head_dim", None)
+        partial_rotary_factor = getattr(config, "partial_rotary_factor", None)
+        dual_chunk_attention_config = getattr(
+            config, "dual_chunk_attention_config", None
+        )
         self.self_attn = Glm4Attention(
-            config, layer_id, quant_config, prefix=add_prefix("self_attn", prefix)
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            head_dim=head_dim,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+            partial_rotary_factor=partial_rotary_factor,
+            prefix=add_prefix("self_attn", prefix),
         )
 
         # MLP
@@ -199,54 +291,125 @@ class Glm4Model(nn.Module):
         config: Glm4Config,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Glm4DecoderLayer,
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ) -> None:
         super().__init__()
         self.config = config
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-            quant_config=quant_config,
-            prefix=add_prefix("embed_tokens", prefix),
-        )
-        self.layers = make_layers(
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.pp_group = get_pp_group()
+
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                enable_tp=not is_dp_attention_enabled(),
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        # Use the provided decoder layer type or default to Glm4DecoderLayer
+        decoder_layer_type = decoder_layer_type or Glm4DecoderLayer
+        self.layers, self.start_layer, self.end_layer = make_layers(
             config.num_hidden_layers,
-            lambda idx, prefix: Glm4DecoderLayer(
-                config=config, layer_id=idx, quant_config=quant_config, prefix=prefix
+            lambda idx, prefix: decoder_layer_type(
+                layer_id=idx,
+                config=config,
+                quant_config=quant_config,
+                prefix=prefix,
+                alt_stream=alt_stream,
             ),
-            prefix="model.layers",
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
+            prefix=add_prefix("layers", prefix),
         )
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
 
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        # For EAGLE3 support
+        self.layers_to_capture = []
 
     def get_input_embeddings(self) -> nn.Embedding:
         return self.embed_tokens
 
-    def dtype(self) -> torch.dtype:
-        return next(self.parameters()).dtype
-
-    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
-        if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
         else:
-            hidden_states = input_embeds
-        residual = None
-        for layer in self.layers:
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        aux_hidden_states = []
+        for i in range(self.start_layer, self.end_layer):
+            if i in self.layers_to_capture:
+                aux_hidden_states.append(
+                    hidden_states + residual if residual is not None else hidden_states
+                )
+            layer = self.layers[i]
             hidden_states, residual = layer(
                 positions,
                 hidden_states,
                 forward_batch,
                 residual,
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            if hidden_states.shape[0] != 0:
+                if residual is None:
+                    hidden_states = self.norm(hidden_states)
+                else:
+                    hidden_states, _ = self.norm(hidden_states, residual)
+
+        if len(aux_hidden_states) == 0:
+            return hidden_states
 
-        return hidden_states
+        return hidden_states, aux_hidden_states
+
+    # If this function is called, it should always initialize KV cache scale
+    # factors (or else raise an exception). Thus, handled exceptions should
+    # make sure to leave KV cache scale factors in a known good (dummy) state
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        for layer_idx, scaling_factor in kv_cache_scales_loader(
+            quantization_param_path,
+            tp_rank,
+            tp_size,
+            self.config.num_hidden_layers,
+            self.config.__class__.model_type,
+        ):
+            if not isinstance(self.layers[layer_idx], nn.Identity):
+                layer_self_attn = self.layers[layer_idx].self_attn
+            if hasattr(layer_self_attn.attn, "k_scale"):
+                layer_self_attn.attn.k_scale = scaling_factor
+                layer_self_attn.attn.v_scale = scaling_factor
+            else:
+                raise RuntimeError(
+                    "Self attention has no KV cache scaling factor attribute!"
+                )
 
 
 class Glm4ForCausalLM(nn.Module):
@@ -255,21 +418,54 @@ class Glm4ForCausalLM(nn.Module):
         config: Glm4Config,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
-    ):
+    ) -> None:
         super().__init__()
-        self.config: Glm4Config = config
+        self.pp_group = get_pp_group()
+        self.config = config
         self.quant_config = quant_config
-        self.model = Glm4Model(config, quant_config, add_prefix("model", prefix))
-        if config.tie_word_embeddings:
-            self.lm_head = self.model.embed_tokens
+        self.model = Glm4Model(
+            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+
+        # handle the lm head on different pp ranks
+        if self.pp_group.is_last_rank:
+            if self.pp_group.world_size == 1 and config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(
+                    config.vocab_size,
+                    config.hidden_size,
+                    quant_config=quant_config,
+                    prefix=add_prefix("lm_head", prefix),
+                )
         else:
-            self.lm_head = ParallelLMHead(
-                config.vocab_size,
-                config.hidden_size,
-                quant_config=quant_config,
-                prefix="lm_head",
-            )
+            # ranks other than the last rank will have a placeholder layer
+            self.lm_head = PPMissingLayer()
+
+        # perform weight tying for PP
+        if self.pp_group.world_size > 1 and config.tie_word_embeddings:
+            if self.pp_group.is_first_rank:
+                self.pp_group.send(
+                    self.model.embed_tokens.weight, dst=self.pp_group.last_rank
+                )
+            else:
+                emb_token_weight = self.pp_group.recv(
+                    size=(config.vocab_size, config.hidden_size),
+                    dtype=next(self.model.parameters()).dtype,
+                    src=self.pp_group.first_rank,
+                )
+                self.lm_head.weight.copy_(emb_token_weight)
+
         self.logits_processor = LogitsProcessor(config)
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+        # For EAGLE3 support
+        self.capture_aux_hidden_states = False
+
+    def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embedding(input_ids)
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.embed_tokens
 
     @torch.no_grad()
     def forward(
@@ -277,34 +473,138 @@ class Glm4ForCausalLM(nn.Module):
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = False,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, forward_batch)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head, forward_batch
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
         )
+        aux_hidden_states = None
+        if self.capture_aux_hidden_states:
+            hidden_states, aux_hidden_states = hidden_states
+
+        if self.pp_group.is_last_rank:
+            if not get_embedding:
+                return self.logits_processor(
+                    input_ids,
+                    hidden_states,
+                    self.lm_head,
+                    forward_batch,
+                    aux_hidden_states,
+                )
+            else:
+                return self.pooler(hidden_states, forward_batch)
+        else:
+            return hidden_states
+
+    @torch.no_grad()
+    def forward_split_prefill(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        split_interval: Tuple[int, int],  # [start, end) 0-based
+        input_embeds: torch.Tensor = None,
+    ):
+        start, end = split_interval
+        # embed
+        if start == 0:
+            if input_embeds is None:
+                forward_batch.hidden_states = self.model.embed_tokens(input_ids)
+            else:
+                forward_batch.hidden_states = input_embeds
+        # decoder layer
+        for i in range(start, end):
+            layer = self.model.layers[i]
+            forward_batch.hidden_states, forward_batch.residual = layer(
+                positions,
+                forward_batch.hidden_states,
+                forward_batch,
+                forward_batch.residual,
+            )
+
+        if end == self.model.config.num_hidden_layers:
+            # norm
+            hidden_states, _ = self.model.norm(
+                forward_batch.hidden_states, forward_batch.residual
+            )
+            forward_batch.hidden_states = hidden_states
+            # logits process
+            result = self.logits_processor(
+                input_ids, forward_batch.hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            result = None
+
+        return result
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
-            # (param_name, weight_name, shard_id)
+            # (param_name, shard_name, shard_id)
             (".qkv_proj", ".q_proj", "q"),
             (".qkv_proj", ".k_proj", "k"),
             (".qkv_proj", ".v_proj", "v"),
-            (".gate_up_proj", ".gate_proj", 0),
             (".gate_up_proj", ".up_proj", 1),
+            (".gate_up_proj", ".gate_proj", 0),
         ]
+
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
-            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
+            if "rotary_emb.inv_freq" in name or "projector" in name:
                 continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                if self.pp_group.world_size > 1 and self.pp_group.is_last_rank:
+                    # Handle pp weight tying here
+                    # find the embed_tokens.weight in the weights
+                    embed_token_weights = next(
+                        filter(lambda x: x[0] == "model.embed_tokens.weight", weights)
+                    )[1]
+                    loaded_weight = embed_token_weights
+                else:
+                    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
+                if 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
+
                 if name in params_dict.keys():
                     param = params_dict[name]
                     weight_loader = getattr(
@@ -312,7 +612,21 @@ class Glm4ForCausalLM(nn.Module):
                     )
                     weight_loader(param, loaded_weight)
                 else:
-                    raise KeyError(f"Parameter '{name}' not found in model.")
+                    logger.warning(f"Parameter {name} not found in params_dict")
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        self.model.load_kv_cache_scales(quantization_param_path)
 
 
 EntryClass = [Glm4ForCausalLM]