Fix the chat template for llava-v1.6-34b & format code (#177)

python/sglang/api.py

@@ -1,4 +1,5 @@
 """Public API"""
+
 import re
 from typing import Callable, List, Optional, Union
 

python/sglang/backend/runtime_endpoint.py

@@ -19,7 +19,9 @@ def __init__(self, base_url, auth_token=None):
         self.base_url = base_url
         self.auth_token = auth_token
 
-        res = http_request(self.base_url + "/get_model_info", auth_token=self.auth_token)
+        res = http_request(
+            self.base_url + "/get_model_info", auth_token=self.auth_token
+        )
         assert res.status_code == 200
         self.model_info = res.json()
 
@@ -37,22 +39,24 @@ def cache_prefix(self, prefix_str: str):
         res = http_request(
             self.base_url + "/generate",
             json={"text": prefix_str, "sampling_params": {"max_new_tokens": 0}},
-            auth_token=self.auth_token
+            auth_token=self.auth_token,
         )
         assert res.status_code == 200
 
     def commit_lazy_operations(self, s: StreamExecutor):
         res = http_request(
             self.base_url + "/generate",
             json={"text": s.text_, "sampling_params": {"max_new_tokens": 0}},
-            auth_token=self.auth_token
+            auth_token=self.auth_token,
         )
         assert res.status_code == 200
 
     def fill_image(self, s: StreamExecutor):
         data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
         self._add_images(s, data)
-        res = http_request(self.base_url + "/generate", json=data, auth_token=self.auth_token)
+        res = http_request(
+            self.base_url + "/generate", json=data, auth_token=self.auth_token
+        )
         assert res.status_code == 200
 
     def generate(
@@ -82,7 +86,9 @@ def generate(
 
         self._add_images(s, data)
 
-        res = http_request(self.base_url + "/generate", json=data, auth_token=self.auth_token)
+        res = http_request(
+            self.base_url + "/generate", json=data, auth_token=self.auth_token
+        )
         obj = res.json()
         comp = obj["text"]
         return comp, obj["meta_info"]
@@ -115,7 +121,12 @@ def generate_stream(
         data["stream"] = True
         self._add_images(s, data)
 
-        response = http_request(self.base_url + "/generate", json=data, stream=True, auth_token=self.auth_token)
+        response = http_request(
+            self.base_url + "/generate",
+            json=data,
+            stream=True,
+            auth_token=self.auth_token,
+        )
         pos = 0
 
         incomplete_text = ""
@@ -145,7 +156,9 @@ def select(
         # Cache common prefix
         data = {"text": s.text_, "sampling_params": {"max_new_tokens": 0}}
         self._add_images(s, data)
-        res = http_request(self.base_url + "/generate", json=data, auth_token=self.auth_token)
+        res = http_request(
+            self.base_url + "/generate", json=data, auth_token=self.auth_token
+        )
         assert res.status_code == 200
         prompt_len = res.json()["meta_info"]["prompt_tokens"]
 
@@ -157,7 +170,9 @@ def select(
             "logprob_start_len": max(prompt_len - 2, 0),
         }
         self._add_images(s, data)
-        res = http_request(self.base_url + "/generate", json=data, auth_token=self.auth_token)
+        res = http_request(
+            self.base_url + "/generate", json=data, auth_token=self.auth_token
+        )
         assert res.status_code == 200
         obj = res.json()
         normalized_prompt_logprob = [
@@ -172,7 +187,7 @@ def concatenate_and_append(self, src_rids: List[str], dst_rid: str):
         res = http_request(
             self.base_url + "/concate_and_append_request",
             json={"src_rids": src_rids, "dst_rid": dst_rid},
-            auth_token=self.auth_token
+            auth_token=self.auth_token,
         )
         assert res.status_code == 200
 

python/sglang/lang/chat_template.py

@@ -116,6 +116,21 @@ def get_chat_template_by_model_path(model_path):
 )
 
 
+register_chat_template(
+    ChatTemplate(
+        name="chatml-llava",
+        default_system_prompt="Answer the questions.",
+        role_prefix_and_suffix={
+            "system": ("<|im_start|>system\n", "\n<|im_end|>\n"),
+            "user": ("<|im_start|>user\n", "\n<|im_end|>\n"),
+            "assistant": ("<|im_start|>assistant\n", "\n<|im_end|>\n"),
+        },
+        style=ChatTemplateStyle.PLAIN,
+        stop_str=("<|im_end|>",),
+        image_token=" <image>\n",
+    )
+)
+
 register_chat_template(
     ChatTemplate(
         name="vicuna_v1.1",
@@ -168,7 +183,7 @@ def get_chat_template_by_model_path(model_path):
 def match_vicuna(model_path: str):
     if "vicuna" in model_path.lower():
         return get_chat_template("vicuna_v1.1")
-    if "llava" in model_path.lower():
+    if "llava-v1.5" in model_path.lower():
         return get_chat_template("vicuna_v1.1")
 
 
@@ -192,6 +207,8 @@ def match_chat_ml(model_path: str):
         return get_chat_template("chatml")
     if "qwen" in model_path and "chat" in model_path:
         return get_chat_template("chatml")
+    if "llava-v1.6-34b" in model_path:
+        return get_chat_template("chatml-llava")
 
 
 @register_chat_template_matching_function

python/sglang/lang/ir.py

@@ -74,9 +74,9 @@ def to_anthropic_kwargs(self):
             )
         return {
             "max_tokens_to_sample": self.max_new_tokens,
-            "stop_sequences": self.stop
-            if isinstance(self.stop, (list, tuple))
-            else [self.stop],
+            "stop_sequences": (
+                self.stop if isinstance(self.stop, (list, tuple)) else [self.stop]
+            ),
             "temperature": self.temperature,
             "top_p": self.top_p,
             "top_k": self.top_k,

python/sglang/lang/tracer.py

@@ -1,4 +1,5 @@
 """Tracing a program."""
+
 import uuid
 from typing import Any, Callable, Dict, List, Optional, Union
 

python/sglang/srt/backend_config.py

@@ -1,6 +1,7 @@
 """
 Backend configurations, may vary with different serving platforms.
 """
+
 from dataclasses import dataclass
 
 

python/sglang/srt/conversation.py

@@ -366,7 +366,8 @@ def generate_chat_conv(
                     if content.type == "text":
                         real_content += content.text
                     elif content.type == "image_url":
-                        real_content += "<image>"
+                        # NOTE: Only works for llava
+                        real_content += "<image>\n"
                         conv.append_image(content.image_url.url)
                 conv.append_message(conv.roles[0], real_content)
         elif msg_role == "assistant":

python/sglang/srt/managers/router/model_rpc.py

@@ -31,6 +31,7 @@
     is_multimodal_model,
     set_random_seed,
 )
+from vllm.logger import _default_handler as vllm_default_handler
 
 logger = logging.getLogger("model_rpc")
 
@@ -50,6 +51,9 @@ def exposed_init_model(
         self.tp_size = server_args.tp_size
         self.schedule_heuristic = server_args.schedule_heuristic
         self.disable_regex_jump_forward = server_args.disable_regex_jump_forward
+        vllm_default_handler.setLevel(
+            level=getattr(logging, server_args.log_level.upper())
+        )
 
         # Init model and tokenizer
         self.model_config = ModelConfig(
@@ -83,9 +87,11 @@ def exposed_init_model(
         self.max_num_running_seq = self.max_total_num_token // 2
         self.max_prefill_num_token = max(
             self.model_config.context_len,
-            self.max_total_num_token // 6
-            if server_args.max_prefill_num_token is None
-            else server_args.max_prefill_num_token,
+            (
+                self.max_total_num_token // 6
+                if server_args.max_prefill_num_token is None
+                else server_args.max_prefill_num_token
+            ),
         )
         self.int_token_logit_bias = torch.tensor(
             get_int_token_logit_bias(self.tokenizer, self.model_config.vocab_size)
@@ -534,7 +540,7 @@ def handle_finished_requests(self, batch: Batch):
                 output_skip_special_tokens.append(
                     req.sampling_params.skip_special_tokens
                 )
-                
+
                 # For the length of input_ids, which will be accumulated during jump-forward.
                 # Use the original length of input_ids to calculate the token usage info.
                 meta_info = {

python/sglang/srt/managers/router/model_runner.py

@@ -112,7 +112,9 @@ def init_flashinfer_args(self, tp_size):
             (self.batch_size,), dtype=torch.int32, device="cuda"
         )
 
-        workspace_buffer = torch.empty(32 * 1024 * 1024, dtype=torch.int8, device="cuda")
+        workspace_buffer = torch.empty(
+            32 * 1024 * 1024, dtype=torch.int8, device="cuda"
+        )
         if (
             self.forward_mode == ForwardMode.PREFILL
             or self.forward_mode == ForwardMode.EXTEND
@@ -121,7 +123,9 @@ def init_flashinfer_args(self, tp_size):
                 (self.batch_size + 1,), dtype=torch.int32, device="cuda"
             )
             self.qo_indptr[1:] = torch.cumsum(self.extend_seq_lens, dim=0)
-            self.prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper(workspace_buffer, "NHD")
+            self.prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper(
+                workspace_buffer, "NHD"
+            )
             self.prefill_wrapper.begin_forward(
                 self.qo_indptr,
                 self.kv_indptr,
@@ -131,7 +135,9 @@ def init_flashinfer_args(self, tp_size):
                 self.model_runner.model_config.num_key_value_heads // tp_size,
             )
         else:
-            self.decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD")
+            self.decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+                workspace_buffer, "NHD"
+            )
             self.decode_wrapper.begin_forward(
                 self.kv_indptr,
                 self.kv_indices,

python/sglang/srt/memory_pool.py

@@ -1,4 +1,5 @@
 """Memory pool."""
+
 import logging
 
 import torch

python/sglang/srt/models/llava.py

@@ -1,4 +1,5 @@
 """Inference-only LLaVa model compatible with HuggingFace weights."""
+
 from typing import List, Optional
 
 import numpy as np
@@ -269,7 +270,6 @@ def load_weights(
             raise ValueError(f"Unexpected select feature: {self.select_feature}")
 
         # load mm_projector
-        # TODO: support TP?
         projector_weights = {
             "model.mm_projector.0": "multi_modal_projector.linear_1",
             "model.mm_projector.2": "multi_modal_projector.linear_2",

python/sglang/srt/models/mistral.py

@@ -1,4 +1,5 @@
 """Inference-only Mistral model."""
+
 from sglang.srt.models.llama2 import LlamaForCausalLM
 
 

python/sglang/srt/models/mixtral.py

@@ -97,14 +97,16 @@ def __init__(
 
         self.experts = nn.ModuleList(
             [
-                MixtralMLP(
-                    self.num_total_experts,
-                    config.hidden_size,
-                    config.intermediate_size,
-                    linear_method=linear_method,
+                (
+                    MixtralMLP(
+                        self.num_total_experts,
+                        config.hidden_size,
+                        config.intermediate_size,
+                        linear_method=linear_method,
+                    )
+                    if idx in self.expert_indicies
+                    else None
                 )
-                if idx in self.expert_indicies
-                else None
                 for idx in range(self.num_total_experts)
             ]
         )

python/sglang/srt/models/yivl.py

@@ -1,4 +1,5 @@
 """Inference-only Yi-VL model."""
+
 import os
 from typing import List, Optional
 

python/sglang/srt/sampling_params.py

@@ -1,4 +1,5 @@
 """Sampling parameters for text generation."""
+
 from typing import List, Optional, Union
 
 _SAMPLING_EPS = 1e-6

python/sglang/srt/server.py

@@ -1,4 +1,5 @@
 """SRT: SGLang Runtime"""
+
 import asyncio
 import json
 import multiprocessing as mp
@@ -493,7 +494,7 @@ def _launch_server():
 
     # Warmup
     try:
-        print("Warmup...", flush=True)
+        # print("Warmup...", flush=True)
         res = requests.post(
             url + "/generate",
             json={
@@ -505,8 +506,8 @@ def _launch_server():
             },
             timeout=60,
         )
-        print(f"Warmup done. model response: {res.json()['text']}")
-        print("=" * 20, "Server is ready", "=" * 20, flush=True)
+        # print(f"Warmup done. model response: {res.json()['text']}")
+        # print("=" * 20, "Server is ready", "=" * 20, flush=True)
     except requests.exceptions.RequestException as e:
         if pipe_finish_writer is not None:
             pipe_finish_writer.send(str(e))

python/sglang/srt/utils.py

@@ -122,7 +122,7 @@ def handle_port_init(
     # first check on server port
     if not check_port(port):
         new_port = alloc_usable_network_port(1, used_list=[port])[0]
-        print(f"Port {port} is not available, using {new_port} instead.")
+        print(f"WARNING: Port {port} is not available. Use {new_port} instead.")
         port = new_port
 
     # then we check on additional ports
@@ -157,8 +157,6 @@ def get_int_token_logit_bias(tokenizer, vocab_size):
         ss = tokenizer.decode([t_id]).strip()
         if not (ss.isdigit() or len(ss) == 0 or t_id == tokenizer.eos_token_id):
             logit_bias[t_id] = -1e5
-        # else:
-        #    print(ss, t_id)
 
     return logit_bias
 

python/sglang/test/test_utils.py

@@ -1,4 +1,5 @@
 """Common utilities for testing and benchmarking"""
+
 import numpy as np
 import requests
 from sglang.backend.openai import OpenAI

python/sglang/utils.py

@@ -22,7 +22,7 @@ def get_available_gpu_memory(gpu_id, distributed=True):
 
     if torch.cuda.current_device() != gpu_id:
         print(
-            f"WARN: current device is not {gpu_id}, but {torch.cuda.current_device()}, ",
+            f"WARNING: current device is not {gpu_id}, but {torch.cuda.current_device()}, ",
             "which may cause useless memory allocation for torch CUDA context.",
         )
 
@@ -95,7 +95,7 @@ def http_request(url, json=None, stream=False, auth_token=None):
             return requests.post(url, json=json, stream=True)
         headers = {
             "Content-Type": "application/json",
-            "Authentication": f"Bearer {auth_token}"
+            "Authentication": f"Bearer {auth_token}",
         }
         return requests.post(url, json=json, stream=True, headers=headers)
     else:

test/lang/test_srt_backend.py

@@ -1,6 +1,7 @@
 """
 python3 -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000
 """
+
 import json
 import unittest