File size: 18,585 Bytes
09b5534
f238f35
7270816
ab0cf4f
09b5534
f238f35
1ea9642
 
51159ea
f238f35
09b5534
f238f35
09b5534
 
ab0cf4f
51159ea
 
f238f35
 
7270816
09b5534
ab0cf4f
 
f238f35
ab0cf4f
 
 
09b5534
ab0cf4f
 
09b5534
 
ab0cf4f
 
 
 
09b5534
f238f35
 
 
09b5534
f238f35
 
09b5534
ab0cf4f
09b5534
 
ab0cf4f
 
 
 
 
 
 
 
1ea9642
7270816
ab0cf4f
1ea9642
 
 
ab0cf4f
09b5534
 
51159ea
 
f238f35
51159ea
09b5534
 
 
f238f35
 
 
51159ea
 
7270816
 
 
 
 
f238f35
 
 
 
 
51159ea
 
f238f35
 
 
 
 
51159ea
 
f238f35
 
 
 
 
51159ea
 
f238f35
 
 
51159ea
09b5534
7270816
09b5534
 
f238f35
 
 
09b5534
51159ea
f238f35
 
 
 
51159ea
 
f238f35
 
 
 
51159ea
 
f238f35
 
 
7270816
51159ea
 
f238f35
 
 
09b5534
 
f238f35
 
09b5534
 
f238f35
 
 
 
 
 
1ea9642
 
f238f35
 
 
7270816
f238f35
 
1ea9642
09b5534
 
f238f35
 
 
 
 
09b5534
 
f238f35
51159ea
f238f35
 
7270816
f238f35
 
 
 
51159ea
 
 
 
f238f35
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7270816
f238f35
 
 
 
 
 
7270816
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f238f35
 
 
 
51159ea
f238f35
 
51159ea
 
 
 
 
7270816
51159ea
 
 
 
 
 
 
 
 
 
 
 
7270816
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ea9642
f238f35
 
 
 
7270816
51159ea
f238f35
 
1ea9642
f238f35
 
51159ea
7270816
 
 
1ea9642
7270816
 
 
1ea9642
 
 
7270816
1ea9642
7270816
 
1ea9642
51159ea
f238f35
1ea9642
51159ea
f238f35
 
7270816
f238f35
1ea9642
 
7270816
f238f35
1ea9642
 
 
f238f35
 
 
 
 
 
 
1ea9642
f238f35
51159ea
 
1ea9642
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51159ea
 
f238f35
 
51159ea
f238f35
7270816
 
 
 
 
 
 
 
 
 
 
51159ea
f238f35
 
1ea9642
f238f35
1ea9642
f238f35
 
7270816
 
 
 
 
 
 
 
1ea9642
7270816
f238f35
 
 
 
 
 
 
 
 
 
 
1ea9642
51159ea
 
 
 
1ea9642
51159ea
f238f35
 
 
09b5534
f238f35
 
 
 
 
 
 
 
 
 
 
 
51159ea
1ea9642
f238f35
1ea9642
 
51159ea
 
 
 
f238f35
51159ea
1ea9642
09b5534
 
f238f35
09b5534
1ea9642
09b5534
 
51159ea
09b5534
1ea9642
09b5534
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
"""
Docker Model Runner - Anthropic API Compatible
Full compatibility with Anthropic Messages API + Interleaved Thinking
Optimized for: 2 vCPU, 16GB RAM
"""
from fastapi import FastAPI, HTTPException, Header, Request
from fastapi.responses import StreamingResponse, HTMLResponse, FileResponse
from fastapi.staticfiles import StaticFiles
from pydantic import BaseModel, Field
from typing import Optional, List, Union, Literal, Any, Dict
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import os
from datetime import datetime
from contextlib import asynccontextmanager
import uuid
import time
import json
import asyncio
import re

# CPU-optimized lightweight models
GENERATOR_MODEL = os.getenv("GENERATOR_MODEL", "distilgpt2")
MODEL_DISPLAY_NAME = os.getenv("MODEL_NAME", "MiniMax-M2")

# Set CPU threading
torch.set_num_threads(2)

# Global model cache
models = {}


def load_models():
    """Pre-load models for faster inference"""
    global models
    print("Loading models for CPU inference...")

    models["tokenizer"] = AutoTokenizer.from_pretrained(GENERATOR_MODEL)
    models["model"] = AutoModelForCausalLM.from_pretrained(GENERATOR_MODEL)
    models["model"].eval()

    if models["tokenizer"].pad_token is None:
        models["tokenizer"].pad_token = models["tokenizer"].eos_token

    print("✅ All models loaded successfully!")


@asynccontextmanager
async def lifespan(app: FastAPI):
    load_models()
    yield
    models.clear()


app = FastAPI(
    title="Model Runner",
    description="Anthropic API Compatible with Interleaved Thinking",
    version="1.0.0",
    lifespan=lifespan,
    docs_url="/api/docs",
    redoc_url="/api/redoc"
)


# ============== Anthropic API Models ==============

class TextBlock(BaseModel):
    type: Literal["text"] = "text"
    text: str


class ThinkingBlock(BaseModel):
    type: Literal["thinking"] = "thinking"
    thinking: str


class SignatureBlock(BaseModel):
    type: Literal["signature"] = "signature"
    signature: str


class ToolUseBlock(BaseModel):
    type: Literal["tool_use"] = "tool_use"
    id: str
    name: str
    input: Dict[str, Any]


class ToolResultContent(BaseModel):
    type: Literal["tool_result"] = "tool_result"
    tool_use_id: str
    content: Union[str, List[TextBlock]]
    is_error: Optional[bool] = False


class ImageSource(BaseModel):
    type: Literal["base64", "url"]
    media_type: Optional[str] = None
    data: Optional[str] = None
    url: Optional[str] = None


class ImageBlock(BaseModel):
    type: Literal["image"] = "image"
    source: ImageSource


ContentBlock = Union[TextBlock, ThinkingBlock, SignatureBlock, ToolUseBlock, ToolResultContent, ImageBlock, str]


class MessageParam(BaseModel):
    role: Literal["user", "assistant"]
    content: Union[str, List[ContentBlock]]


class ToolInputSchema(BaseModel):
    type: str = "object"
    properties: Optional[Dict[str, Any]] = None
    required: Optional[List[str]] = None


class Tool(BaseModel):
    name: str
    description: str
    input_schema: ToolInputSchema


class ToolChoice(BaseModel):
    type: Literal["auto", "any", "tool"] = "auto"
    name: Optional[str] = None
    disable_parallel_tool_use: Optional[bool] = False


class ThinkingConfig(BaseModel):
    type: Literal["enabled", "disabled"] = "disabled"
    budget_tokens: Optional[int] = None


class Metadata(BaseModel):
    user_id: Optional[str] = None


class AnthropicRequest(BaseModel):
    model: str = "MiniMax-M2"
    messages: List[MessageParam]
    max_tokens: int = 1024
    temperature: Optional[float] = Field(default=1.0, gt=0.0, le=1.0)
    top_p: Optional[float] = Field(default=1.0, gt=0.0, le=1.0)
    top_k: Optional[int] = None
    stop_sequences: Optional[List[str]] = None
    stream: Optional[bool] = False
    system: Optional[Union[str, List[TextBlock]]] = None
    tools: Optional[List[Tool]] = None
    tool_choice: Optional[Union[ToolChoice, Dict[str, Any]]] = None
    metadata: Optional[Metadata] = None
    thinking: Optional[ThinkingConfig] = None
    service_tier: Optional[str] = None


class Usage(BaseModel):
    input_tokens: int
    output_tokens: int
    cache_creation_input_tokens: Optional[int] = 0
    cache_read_input_tokens: Optional[int] = 0


class AnthropicResponse(BaseModel):
    id: str
    type: Literal["message"] = "message"
    role: Literal["assistant"] = "assistant"
    content: List[Union[TextBlock, ThinkingBlock, SignatureBlock, ToolUseBlock]]
    model: str
    stop_reason: Optional[Literal["end_turn", "max_tokens", "stop_sequence", "tool_use"]] = "end_turn"
    stop_sequence: Optional[str] = None
    usage: Usage


# ============== Helper Functions ==============

def extract_text_from_content(content: Union[str, List[ContentBlock]]) -> str:
    if isinstance(content, str):
        return content
    texts = []
    for block in content:
        if isinstance(block, str):
            texts.append(block)
        elif hasattr(block, 'text'):
            texts.append(block.text)
        elif hasattr(block, 'thinking'):
            texts.append(block.thinking)
        elif isinstance(block, dict):
            if block.get('type') == 'text':
                texts.append(block.get('text', ''))
            elif block.get('type') == 'thinking':
                texts.append(block.get('thinking', ''))
    return " ".join(texts)


def format_system_prompt(system: Optional[Union[str, List[TextBlock]]]) -> str:
    if system is None:
        return ""
    if isinstance(system, str):
        return system
    return " ".join([block.text for block in system if hasattr(block, 'text')])


def format_messages_to_prompt(messages: List[MessageParam], system: Optional[Union[str, List[TextBlock]]] = None, include_thinking: bool = False) -> str:
    prompt_parts = []
    system_text = format_system_prompt(system)
    if system_text:
        prompt_parts.append(f"System: {system_text}\n\n")
    for msg in messages:
        role = msg.role
        content = msg.content
        if isinstance(content, list):
            for block in content:
                if isinstance(block, dict):
                    block_type = block.get('type', 'text')
                    if block_type == 'thinking' and include_thinking:
                        prompt_parts.append(f"<thinking>{block.get('thinking', '')}</thinking>\n")
                    elif block_type == 'text':
                        if role == "user":
                            prompt_parts.append(f"Human: {block.get('text', '')}\n\n")
                        else:
                            prompt_parts.append(f"Assistant: {block.get('text', '')}\n\n")
                elif hasattr(block, 'type'):
                    if block.type == 'thinking' and include_thinking:
                        prompt_parts.append(f"<thinking>{block.thinking}</thinking>\n")
                    elif block.type == 'text':
                        if role == "user":
                            prompt_parts.append(f"Human: {block.text}\n\n")
                        else:
                            prompt_parts.append(f"Assistant: {block.text}\n\n")
        else:
            content_text = content if isinstance(content, str) else extract_text_from_content(content)
            if role == "user":
                prompt_parts.append(f"Human: {content_text}\n\n")
            elif role == "assistant":
                prompt_parts.append(f"Assistant: {content_text}\n\n")
    prompt_parts.append("Assistant:")
    return "".join(prompt_parts)


def generate_text(prompt: str, max_tokens: int, temperature: float, top_p: float) -> tuple:
    tokenizer = models["tokenizer"]
    model = models["model"]
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
    input_tokens = inputs["input_ids"].shape[1]
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=min(max_tokens, 256),
            temperature=temperature if temperature > 0 else 1.0,
            top_p=top_p,
            do_sample=temperature > 0,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id
        )
    generated_tokens = outputs[0][input_tokens:]
    output_tokens = len(generated_tokens)
    generated_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
    return generated_text.strip(), input_tokens, output_tokens


def generate_thinking(prompt: str, budget_tokens: int = 100) -> tuple:
    tokenizer = models["tokenizer"]
    model = models["model"]
    thinking_prompt = f"{prompt}\n\nLet me think through this step by step:\n"
    inputs = tokenizer(thinking_prompt, return_tensors="pt", truncation=True, max_length=512)
    input_tokens = inputs["input_ids"].shape[1]
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=min(budget_tokens, 128),
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id
        )
    generated_tokens = outputs[0][input_tokens:]
    thinking_tokens = len(generated_tokens)
    thinking_text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
    return thinking_text.strip(), thinking_tokens


async def generate_stream_with_thinking(prompt: str, max_tokens: int, temperature: float, top_p: float, message_id: str, model_name: str, thinking_enabled: bool = False, thinking_budget: int = 100):
    tokenizer = models["tokenizer"]
    model = models["model"]
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
    input_tokens = inputs["input_ids"].shape[1]
    total_output_tokens = 0

    message_start = {
        "type": "message_start",
        "message": {"id": message_id, "type": "message", "role": "assistant", "content": [], "model": model_name, "stop_reason": None, "stop_sequence": None, "usage": {"input_tokens": input_tokens, "output_tokens": 0}}
    }
    yield f"event: message_start\ndata: {json.dumps(message_start)}\n\n"

    content_index = 0

    if thinking_enabled:
        thinking_block_start = {"type": "content_block_start", "index": content_index, "content_block": {"type": "thinking", "thinking": ""}}
        yield f"event: content_block_start\ndata: {json.dumps(thinking_block_start)}\n\n"
        thinking_text, thinking_tokens = generate_thinking(prompt, thinking_budget)
        total_output_tokens += thinking_tokens
        for i in range(0, len(thinking_text), 10):
            chunk = thinking_text[i:i+10]
            yield f"event: content_block_delta\ndata: {json.dumps({'type': 'content_block_delta', 'index': content_index, 'delta': {'type': 'thinking_delta', 'thinking': chunk}})}\n\n"
            await asyncio.sleep(0.01)
        yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': content_index})}\n\n"
        content_index += 1

    yield f"event: content_block_start\ndata: {json.dumps({'type': 'content_block_start', 'index': content_index, 'content_block': {'type': 'text', 'text': ''}})}\n\n"

    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=min(max_tokens, 256), temperature=temperature if temperature > 0 else 1.0, top_p=top_p, do_sample=temperature > 0, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id)

    generated_tokens = outputs[0][input_tokens:]
    generated_text = tokenizer.decode(generated_tokens, skip_special_tokens=True).strip()
    total_output_tokens += len(generated_tokens)

    for i in range(0, len(generated_text), 5):
        yield f"event: content_block_delta\ndata: {json.dumps({'type': 'content_block_delta', 'index': content_index, 'delta': {'type': 'text_delta', 'text': generated_text[i:i+5]}})}\n\n"
        await asyncio.sleep(0.01)

    yield f"event: content_block_stop\ndata: {json.dumps({'type': 'content_block_stop', 'index': content_index})}\n\n"
    yield f"event: message_delta\ndata: {json.dumps({'type': 'message_delta', 'delta': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'usage': {'output_tokens': total_output_tokens}})}\n\n"
    yield f"event: message_stop\ndata: {json.dumps({'type': 'message_stop'})}\n\n"


def handle_tool_call(tools: List[Tool], messages: List[MessageParam], generated_text: str) -> Optional[ToolUseBlock]:
    if not tools:
        return None
    for tool in tools:
        if tool.name.lower() in generated_text.lower():
            return ToolUseBlock(type="tool_use", id=f"toolu_{uuid.uuid4().hex[:24]}", name=tool.name, input={})
    return None


# ============== Frontend ==============

@app.get("/", response_class=HTMLResponse)
async def home():
    """Serve the minimal centered frontend"""
    try:
        with open("/app/static/index.html", "r") as f:
            return HTMLResponse(content=f.read())
    except:
        return HTMLResponse(content="""
<!DOCTYPE html>
<html><head><meta charset="UTF-8"><title>Model Runner</title>
<style>*{margin:0;padding:0}body{min-height:100vh;background:#000;display:flex;justify-content:center;align-items:center}
.logo{width:200px;height:200px;animation:float 3s ease-in-out infinite}
@keyframes float{0%,100%{transform:translateY(0)}50%{transform:translateY(-10px)}}</style></head>
<body><div class="logo"><svg viewBox="0 0 200 200" fill="none">
<defs><linearGradient id="r" x1="0%" y1="100%" x2="100%" y2="0%">
<stop offset="0%" stop-color="#ff0080"/><stop offset="25%" stop-color="#ff4d00"/>
<stop offset="50%" stop-color="#ffcc00"/><stop offset="75%" stop-color="#00ff88"/>
<stop offset="100%" stop-color="#00ccff"/></linearGradient></defs>
<path d="M100 20 L180 160 L20 160 Z" stroke="url(#r)" stroke-width="12" stroke-linecap="round" fill="none"/>
<path d="M100 70 L130 130 L70 130 Z" stroke="url(#r)" stroke-width="8" stroke-linecap="round" fill="none"/>
<line x1="80" y1="115" x2="120" y2="115" stroke="url(#r)" stroke-width="6" stroke-linecap="round"/>
</svg></div></body></html>
        """)


# ============== Anthropic API Endpoints ==============

@app.post("/v1/messages")
async def create_message(request: AnthropicRequest):
    try:
        message_id = f"msg_{uuid.uuid4().hex[:24]}"
        thinking_enabled = False
        thinking_budget = 100
        if request.thinking:
            if isinstance(request.thinking, dict):
                thinking_enabled = request.thinking.get('type') == 'enabled'
                thinking_budget = request.thinking.get('budget_tokens', 100)
            else:
                thinking_enabled = request.thinking.type == 'enabled'
                thinking_budget = request.thinking.budget_tokens or 100

        prompt = format_messages_to_prompt(request.messages, request.system, include_thinking=thinking_enabled)

        if request.stream:
            return StreamingResponse(
                generate_stream_with_thinking(prompt, request.max_tokens, request.temperature or 1.0, request.top_p or 1.0, message_id, request.model, thinking_enabled, thinking_budget),
                media_type="text/event-stream",
                headers={"Cache-Control": "no-cache", "Connection": "keep-alive", "X-Accel-Buffering": "no"}
            )

        content_blocks = []
        total_output_tokens = 0

        if thinking_enabled:
            thinking_text, thinking_tokens = generate_thinking(prompt, thinking_budget)
            total_output_tokens += thinking_tokens
            content_blocks.append(ThinkingBlock(type="thinking", thinking=thinking_text))

        generated_text, input_tokens, output_tokens = generate_text(prompt, request.max_tokens, request.temperature or 1.0, request.top_p or 1.0)
        total_output_tokens += output_tokens

        tool_use = handle_tool_call(request.tools, request.messages, generated_text) if request.tools else None

        if tool_use:
            content_blocks.append(TextBlock(type="text", text=generated_text))
            content_blocks.append(tool_use)
            stop_reason = "tool_use"
        else:
            content_blocks.append(TextBlock(type="text", text=generated_text))
            stop_reason = "end_turn"

        return AnthropicResponse(id=message_id, content=content_blocks, model=request.model, stop_reason=stop_reason, usage=Usage(input_tokens=input_tokens, output_tokens=total_output_tokens))
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))


# ============== OpenAI Compatible ==============

class ChatMessage(BaseModel):
    role: str
    content: str


class ChatCompletionRequest(BaseModel):
    model: str = "distilgpt2"
    messages: List[ChatMessage]
    max_tokens: Optional[int] = 1024
    temperature: Optional[float] = 0.7
    top_p: Optional[float] = 1.0
    stream: Optional[bool] = False


@app.post("/v1/chat/completions")
async def chat_completions(request: ChatCompletionRequest):
    try:
        anthropic_messages = [MessageParam(role=msg.role if msg.role in ["user", "assistant"] else "user", content=msg.content) for msg in request.messages if msg.role in ["user", "assistant"]]
        prompt = format_messages_to_prompt(anthropic_messages)
        generated_text, input_tokens, output_tokens = generate_text(prompt, request.max_tokens or 1024, request.temperature or 0.7, request.top_p or 1.0)
        return {"id": f"chatcmpl-{uuid.uuid4().hex[:24]}", "object": "chat.completion", "created": int(time.time()), "model": request.model, "choices": [{"index": 0, "message": {"role": "assistant", "content": generated_text}, "finish_reason": "stop"}], "usage": {"prompt_tokens": input_tokens, "completion_tokens": output_tokens, "total_tokens": input_tokens + output_tokens}}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))


@app.get("/v1/models")
async def list_models():
    return {"object": "list", "data": [{"id": "MiniMax-M2", "object": "model", "created": int(time.time()), "owned_by": "local"}, {"id": "MiniMax-M2-Stable", "object": "model", "created": int(time.time()), "owned_by": "local"}, {"id": GENERATOR_MODEL, "object": "model", "created": int(time.time()), "owned_by": "local"}]}


@app.get("/health")
async def health():
    return {"status": "healthy", "timestamp": datetime.utcnow().isoformat(), "models_loaded": len(models) > 0}


@app.get("/info")
async def info():
    return {"name": "Model Runner", "version": "1.1.0", "api_compatibility": ["anthropic", "openai"], "interleaved_thinking": True}


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)