sglang 0.1.13__py3-none-any.whl → 0.1.15__py3-none-any.whl

sglang/srt/managers/tokenizer_manager.py

@@ -1,6 +1,7 @@
 import asyncio
 import concurrent.futures
 import dataclasses
+import logging
 import multiprocessing as mp
 import os
 from typing import List
@@ -10,6 +11,7 @@ import transformers
 import uvloop
 import zmq
 import zmq.asyncio
+
 from sglang.srt.hf_transformers_utils import (
     get_config,
     get_context_length,
@@ -30,13 +32,14 @@ from sglang.srt.utils import get_exception_traceback, is_multimodal_model, load_
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
+logger = logging.getLogger(__name__)
+
 
 @dataclasses.dataclass
 class ReqState:
     out_list: List
     finished: bool
     event: asyncio.Event
-    lock: asyncio.Lock
 
 
 global global_processor
@@ -174,18 +177,26 @@ class TokenizerManager:
                 sampling_params=sampling_params,
                 return_logprob=obj.return_logprob,
                 logprob_start_len=obj.logprob_start_len,
+                top_logprobs_num=obj.top_logprobs_num,
                 stream=obj.stream,
             )
             self.send_to_router.send_pyobj(tokenized_obj)
 
-            lock = asyncio.Lock()
             event = asyncio.Event()
-            state = ReqState([], False, event, lock)
+            state = ReqState([], False, event)
             self.rid_to_state[rid] = state
 
             while True:
                 await event.wait()
-                yield state.out_list[-1]
+                out = self.convert_logprob_style(state.out_list[-1],
+                                                 obj.return_logprob,
+                                                 obj.top_logprobs_num,
+                                                 obj.return_text_in_logprobs)
+
+                if self.server_args.log_requests and state.finished:
+                    logger.info(f"in={obj.text}, out={out}")
+
+                yield out
                 state.out_list = []
                 if state.finished:
                     del self.rid_to_state[rid]
@@ -217,13 +228,13 @@ class TokenizerManager:
                     sampling_params=sampling_params,
                     return_logprob=obj.return_logprob[i],
                     logprob_start_len=obj.logprob_start_len[i],
+                    top_logprobs_num=obj.top_logprobs_num[i],
                     stream=obj.stream,
                 )
                 self.send_to_router.send_pyobj(tokenized_obj)
 
-                lock = asyncio.Lock()
                 event = asyncio.Event()
-                state = ReqState([], False, event, lock)
+                state = ReqState([], False, event)
                 self.rid_to_state[rid] = state
 
             output_list = []
@@ -231,16 +242,16 @@ class TokenizerManager:
                 rid = obj.rid[i]
                 state = self.rid_to_state[rid]
                 await state.event.wait()
-                output_list.append(state.out_list[-1])
+                output_list.append(
+                    self.convert_logprob_style(state.out_list[-1],
+                                               obj.return_logprob[i],
+                                               obj.top_logprobs_num[i],
+                                               obj.return_text_in_logprobs))
                 assert state.finished
                 del self.rid_to_state[rid]
 
             yield output_list
 
-    async def detokenize(self, obj: DetokenizeReqInput):
-        token_texts = self.tokenizer.convert_ids_to_tokens(obj.input_ids)
-        return [t.decode() if isinstance(t, bytes) else t for t in token_texts]
-
     async def flush_cache(self):
         flush_cache_req = FlushCacheReq()
         self.send_to_router.send_pyobj(flush_cache_req)
@@ -267,3 +278,37 @@ class TokenizerManager:
                     state.event.set()
             else:
                 raise ValueError(f"Invalid object: {recv_obj}")
+
+    def convert_logprob_style(self, ret, return_logprob, top_logprobs_num, return_text_in_logprobs):
+        if return_logprob:
+            ret["meta_info"]["prefill_token_logprobs"] = self.detokenize_logprob_tokens(
+                ret["meta_info"]["prefill_token_logprobs"], return_text_in_logprobs
+            )
+            ret["meta_info"]["decode_token_logprobs"] = self.detokenize_logprob_tokens(
+                ret["meta_info"]["decode_token_logprobs"], return_text_in_logprobs
+            )
+        if top_logprobs_num > 0:
+            ret["meta_info"]["prefill_top_logprobs"] = self.detokenize_top_logprobs_tokens(
+                ret["meta_info"]["prefill_top_logprobs"], return_text_in_logprobs
+            )
+            ret["meta_info"]["decode_top_logprobs"] = self.detokenize_top_logprobs_tokens(
+                ret["meta_info"]["decode_top_logprobs"], return_text_in_logprobs
+            )
+        return ret
+
+    def detokenize_logprob_tokens(self, token_logprobs, decode_to_text):
+        if not decode_to_text:
+            return [(logprob, token_id, None) for logprob, token_id in token_logprobs]
+
+        token_ids = [tid for _, tid in token_logprobs]
+        token_texts = self.tokenizer.batch_decode(token_ids)
+        return [
+            (logprob, token_id, token_text)
+            for (logprob, token_id), token_text, in zip(token_logprobs, token_texts)
+        ]
+
+    def detokenize_top_logprobs_tokens(self, top_logprobs, decode_to_text):
+        for i, t in enumerate(top_logprobs):
+            if t:
+                top_logprobs[i] = self.detokenize_logprob_tokens(t, decode_to_text)
+        return top_logprobs

sglang/srt/memory_pool.py

@@ -31,9 +31,6 @@ class ReqToTokenPool:
             self.can_use_mem_size += free_index.shape[0]
         self.mem_state[free_index] = 1
 
-        # if self.can_use_mem_size == len(self.mem_state):
-        #     print(f"ReqToTokenPool: freed all. size = {self.can_use_mem_size}.")
-
     def clear(self):
         self.mem_state.fill_(1)
         self.can_use_mem_size = len(self.mem_state)
@@ -42,7 +39,7 @@ class ReqToTokenPool:
 class TokenToKVPool:
     def __init__(self, size, dtype, head_num, head_dim, layer_num):
         self.mem_state = torch.zeros((size,), dtype=torch.int16, device="cuda")
-        self.alloc_ct = 0
+        self.total_ref_ct = 0
 
         # [size, key/value, head_num, head_dim] for each layer
         self.kv_data = [
@@ -83,9 +80,6 @@ class TokenToKVPool:
         self.add_refs(select_index)
         return select_index.to(torch.int32), start_loc, start_loc + need_size
 
-    def free(self, free_index):
-        return self.decrease_refs(free_index)
-
     def used_size(self):
         return len(torch.nonzero(self.mem_state).squeeze(1))
 
@@ -93,20 +87,17 @@ class TokenToKVPool:
         return torch.sum(self.mem_state == 0).item()
 
     def add_refs(self, token_index: torch.Tensor):
-        self.alloc_ct += len(token_index)
+        self.total_ref_ct += len(token_index)
         self.mem_state[token_index] += 1
 
-    def decrease_refs(self, token_index: torch.Tensor):
-        self.alloc_ct -= len(token_index)
+    def dec_refs(self, token_index: torch.Tensor):
+        self.total_ref_ct -= len(token_index)
         self.mem_state[token_index] -= 1
 
         num_freed = torch.sum(self.mem_state[token_index] == 0)
 
-        # if self.alloc_ct == 0:
-        #     print(f"TokenToKVPool: freed all. size = {len(self.mem_state)}.")
-
         return num_freed
 
     def clear(self):
         self.mem_state.fill_(0)
-        self.alloc_ct = 0
+        self.total_ref_ct = 0

sglang/srt/model_config.py

@@ -29,6 +29,13 @@ class ModelConfig:
         )
         self.num_attention_heads = self.hf_config.num_attention_heads
         self.num_key_value_heads = getattr(self.hf_config, "num_key_value_heads", None)
+
+        # for Dbrx and MPT models
+        if self.hf_config.model_type in ["dbrx", "mpt"]:
+            self.num_key_value_heads = getattr(
+                self.hf_config.attn_config, "kv_n_heads", None
+            )
+
         if self.num_key_value_heads is None:
             self.num_key_value_heads = self.num_attention_heads
         self.hidden_size = self.hf_config.hidden_size

sglang/srt/models/commandr.py

@@ -0,0 +1,376 @@
+# coding=utf-8
+# Copyright 2024 Cohere and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This file is based on the LLama model definition file in transformers
+"""PyTorch Cohere model."""
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.parameter import Parameter
+from transformers import PretrainedConfig
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.linear import (
+    MergedColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.utils import set_weight_attrs
+from sglang.srt.weight_utils import (
+    default_weight_loader,
+    hf_model_weights_iterator,
+)
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+
+
+@torch.compile
+def layer_norm_func(hidden_states, weight, variance_epsilon):
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(torch.float32)
+    mean = hidden_states.mean(-1, keepdim=True)
+    variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
+    hidden_states = (hidden_states - mean) * torch.rsqrt(variance + variance_epsilon)
+    hidden_states = weight.to(torch.float32) * hidden_states
+    return hidden_states.to(input_dtype)
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, param_shape=None, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(param_shape))
+        self.variance_epsilon = eps
+        set_weight_attrs(self.weight, {"weight_loader": self.weight_loader})
+
+    def forward(self, hidden_states, residuals=None):
+        hidden_states = layer_norm_func(
+            hidden_states, self.weight, self.variance_epsilon
+        )
+        return hidden_states, residuals
+
+    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
+        tp_rank = get_tensor_model_parallel_rank()
+        shard_dim = 0 if param.dim() != 1 else None
+        param_data = param.data
+        if shard_dim is not None:
+            shard_size = param_data.shape[shard_dim]
+            start_idx = tp_rank * shard_size
+            loaded_weight = loaded_weight.narrow(shard_dim, start_idx, shard_size)
+        assert param_data.shape == loaded_weight.shape
+        param_data.copy_(loaded_weight)
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaMLP Llama->Cohere
+class CohereMLP(nn.Module):
+    def __init__(
+        self,
+        config,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_up_proj = MergedColumnParallelLinear(
+            self.hidden_size,
+            [self.intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.down_proj = RowParallelLinear(
+            self.intermediate_size,
+            self.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class CohereAttention(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        tp_size = get_tensor_model_parallel_world_size()
+        self.config = config
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.total_num_heads = config.num_attention_heads
+        self.num_heads = self.total_num_heads // tp_size
+        self.head_dim = self.hidden_size // self.total_num_heads
+        self.total_num_kv_heads = config.num_key_value_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.
+            assert self.total_num_kv_heads % 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)
+        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.max_position_embeddings = getattr(
+            config, "model_max_length", None
+        ) or getattr(config, "max_position_embeddings", 8192)
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = getattr(config, "rope_scaling", None)
+        self.use_qk_norm = getattr(config, "use_qk_norm", False)
+        self.qkv_proj = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            self.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position_embeddings,
+            base=self.rope_theta,
+            rope_scaling=self.rope_scaling,
+            is_neox_style=False,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+        )
+        if self.use_qk_norm:
+            self.q_norm = LayerNorm(
+                param_shape=(self.num_heads, self.head_dim), eps=config.layer_norm_eps
+            )
+            self.k_norm = LayerNorm(
+                param_shape=(self.num_kv_heads, self.head_dim),
+                eps=config.layer_norm_eps,
+            )
+
+    def _apply_qk_norm(self, q, k):
+        q = q.view(*q.shape[:-1], -1, self.head_dim)
+        k = k.view(*k.shape[:-1], -1, self.head_dim)
+        q, _ = self.q_norm(q)
+        k, _ = self.k_norm(k)
+        q = q.view(*q.shape[:-2], -1)
+        k = k.view(*k.shape[:-2], -1)
+        return q, k
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        if self.use_qk_norm:
+            q, k = self._apply_qk_norm(q, k)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class CohereDecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = CohereAttention(
+            config, layer_id=layer_id, quant_config=quant_config
+        )
+
+        self.mlp = CohereMLP(config, quant_config=quant_config)
+        self.input_layernorm = LayerNorm(
+            param_shape=(config.hidden_size), eps=config.layer_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        residual = hidden_states
+        hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states_attention = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            input_metadata=input_metadata,
+        )
+        hidden_states_mlp = self.mlp(hidden_states)
+        # Add everything together
+        hidden_states = residual + hidden_states_attention + hidden_states_mlp
+
+        return hidden_states, residual
+
+
+class CohereModel(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size, config.hidden_size
+        )
+        self.layers = nn.ModuleList(
+            [
+                CohereDecoderLayer(config, i, quant_config=quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = LayerNorm(
+            param_shape=(config.hidden_size), eps=config.layer_norm_eps
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.embed_tokens(input_ids)
+        residual = None
+        for i in range(len(self.layers)):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                input_metadata,
+                residual,
+            )
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class CohereForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.logits_processor = LogitsProcessor(config)
+        self.model = CohereModel(config, quant_config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            input_metadata,
+        )
+        return self.logits_processor(
+            input_ids, hidden_states, self.model.embed_tokens.weight, input_metadata
+        )
+
+    def load_weights(
+        self,
+        model_name_or_path: str,
+        cache_dir: Optional[str] = None,
+        load_format: str = "auto",
+        revision: Optional[str] = None,
+    ):
+        stacked_params_mapping = [
+            # (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),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params = set()
+        for name, loaded_weight in hf_model_weights_iterator(
+            model_name_or_path, cache_dir, load_format, revision
+        ):
+            for param_name, shard_name, shard_id in stacked_params_mapping:
+                if shard_name not in name:
+                    continue
+                name = name.replace(shard_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:
+                # lm_head is not used in vllm as it is tied with embed_token.
+                # To prevent errors, skip loading lm_head.weight.
+                if "lm_head.weight" in name:
+                    continue
+                # 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)
+            loaded_params.add(name)
+
+
+EntryClass = CohereForCausalLM