File size: 13,675 Bytes
b9607cb
09729dc
e452044
 
09729dc
 
 
 
 
 
 
 
 
0467b75
09729dc
e452044
09729dc
b9607cb
0467b75
 
6a0c90b
0467b75
 
 
 
6a0c90b
0467b75
 
 
 
 
33d45b3
ab1c2e4
36923a0
 
e452044
36923a0
33d45b3
0467b75
 
 
e452044
0467b75
ab1c2e4
0467b75
 
 
 
 
ab1c2e4
0467b75
ab1c2e4
0467b75
 
ab1c2e4
0467b75
ab1c2e4
e452044
0467b75
 
 
1eff8e4
c61ba72
ab1c2e4
0467b75
0711c19
 
 
 
 
 
0467b75
 
 
 
 
 
 
 
 
 
 
 
 
ab1c2e4
c61ba72
 
 
ab1c2e4
09811a8
c61ba72
 
09811a8
 
c61ba72
09811a8
c61ba72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
09811a8
 
 
c61ba72
 
 
 
 
 
 
 
 
 
 
 
 
55475f8
c61ba72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
55475f8
c61ba72
 
 
 
 
 
 
 
 
 
 
ab1c2e4
0467b75
ab1c2e4
0467b75
 
 
 
ab1c2e4
0467b75
 
 
 
c61ba72
0467b75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ab1c2e4
0711c19
0467b75
0711c19
0467b75
 
 
 
 
ab1c2e4
0467b75
 
 
 
ab1c2e4
0467b75
 
 
 
 
 
 
 
ab1c2e4
0467b75
 
 
 
 
0711c19
0467b75
 
 
 
e452044
0467b75
e452044
0467b75
 
09729dc
0467b75
09729dc
0467b75
 
 
 
 
 
 
b9607cb
ab1c2e4
0467b75
 
 
 
 
 
 
09729dc
0467b75
 
 
 
09729dc
0467b75
 
 
 
 
 
 
 
 
 
 
 
ab1c2e4
33d45b3
ab1c2e4
 
0467b75
 
 
 
 
36923a0
0467b75
 
 
 
 
 
 
 
 
ab1c2e4
0467b75
 
 
 
ab1c2e4
0467b75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
09729dc
 
ab1c2e4
0467b75
 
 
 
 
 
 
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
from typing import Dict, List, Any
import torch
import json
import os
import logging

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

class EndpointHandler:
    def __init__(self, path: str = ""):
        """
        Initialize handler using CTransformers format for memory efficiency
        """
        logger.info(f"Loading model from {path}")
        
        try:
            # Use CTransformers format for lower memory usage
            ctransformers_path = os.path.join(path, "models", "ctransformers")
            
            if not os.path.exists(ctransformers_path):
                logger.warning(f"CTransformers path not found: {ctransformers_path}")
                logger.info("Falling back to HuggingFace format")
                ctransformers_path = path
            
            logger.info(f"Using model path: {ctransformers_path}")
            
            # Load components using the working handler approach
            self.tokenizer = self._load_tokenizer(ctransformers_path)
            self.model = self._load_model(ctransformers_path)
            
            logger.info("Model and tokenizer loaded successfully")
            
        except Exception as e:
            logger.error(f"Failed to initialize: {str(e)}")
            raise e
    
    def _load_tokenizer(self, model_path: str):
        """Load tokenizer using AutoTokenizer"""
        logger.info("Loading tokenizer...")
        
        from transformers import AutoTokenizer
        
        tokenizer = AutoTokenizer.from_pretrained(
            model_path,
            trust_remote_code=True,
            use_fast=True,
        )
        
        # Ensure special tokens are set
        if not hasattr(tokenizer, 'pad_token') or tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token
            tokenizer.pad_token_id = tokenizer.eos_token_id
        
        logger.info("Tokenizer loaded successfully")
        return tokenizer
    
    def _load_model(self, model_path: str):
        """Load model using AutoModelForCausalLM with memory optimization"""
        logger.info("Loading model with memory optimization...")
        
        from transformers import AutoModelForCausalLM
        
        # Check GPU availability
        if torch.cuda.is_available():
            logger.info(f"CUDA available: {torch.cuda.get_device_name()}")
            logger.info(f"GPU memory total: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f}GB")
        else:
            logger.warning("CUDA not available, using CPU")
        
        # Memory optimization settings
        device_map = "auto" if torch.cuda.is_available() else None
        gpu_mem = os.environ.get("GPU_MAX_MEM", "10GiB")  # Conservative for 12GB limit
        cpu_mem = os.environ.get("CPU_MAX_MEM", "24GiB")
        max_memory = {0: gpu_mem, "cpu": cpu_mem} if torch.cuda.is_available() else None
        
        # Offload folder for memory management
        offload_folder = os.environ.get("OFFLOAD_FOLDER", "/tmp/hf-offload")
        try:
            os.makedirs(offload_folder, exist_ok=True)
        except OSError:
            offload_folder = "/tmp/hf-offload"
            os.makedirs(offload_folder, exist_ok=True)
        
        # Try to load with quantization first, fallback without if it fails
        model = None
        quantization_config = None
        
        # Attempt 1: Try with 8-bit quantization (if bitsandbytes is available)
        if torch.cuda.is_available():
            try:
                # Check if bitsandbytes is available
                import bitsandbytes
                from transformers import BitsAndBytesConfig
                logger.info("bitsandbytes available, attempting 8-bit quantization...")
                
                bnb_config = BitsAndBytesConfig(load_in_8bit=True)
                
                model = AutoModelForCausalLM.from_pretrained(
                    model_path,
                    trust_remote_code=True,
                    device_map=device_map,
                    quantization_config=bnb_config,
                    low_cpu_mem_usage=True,
                    offload_folder=offload_folder,
                    max_memory=max_memory,
                )
                logger.info("Successfully loaded with 8-bit quantization")
                quantization_config = "8-bit"
                
            except ImportError as e:
                logger.info(f"bitsandbytes not available ({str(e)}), skipping quantization...")
                model = None
            except Exception as e:
                logger.warning(f"8-bit quantization failed: {str(e)}")
                logger.info("Falling back to FP16 without quantization...")
                model = None
        
        # Attempt 2: Fallback to FP16 without quantization
        if model is None:
            try:
                model = AutoModelForCausalLM.from_pretrained(
                    model_path,
                    trust_remote_code=True,
                    device_map=device_map,
                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
                    quantization_config=None,  # Disable model's built-in quantization
                    low_cpu_mem_usage=True,
                    offload_folder=offload_folder if device_map == "auto" else None,
                    max_memory=max_memory,
                )
                logger.info("Successfully loaded with FP16 (no quantization)")
                quantization_config = "fp16"
                
            except Exception as e:
                logger.warning(f"FP16 loading failed: {str(e)}")
                logger.info("Falling back to FP32 CPU loading...")
                model = None
        
        # Attempt 3: Final fallback to CPU FP32
        if model is None:
            try:
                model = AutoModelForCausalLM.from_pretrained(
                    model_path,
                    trust_remote_code=True,
                    torch_dtype=torch.float32,
                    quantization_config=None,  # Disable model's built-in quantization
                    low_cpu_mem_usage=True,
                )
                logger.info("Successfully loaded with FP32 on CPU")
                quantization_config = "fp32_cpu"
                
            except Exception as e:
                logger.error(f"All loading attempts failed: {str(e)}")
                raise e
        
        if model is None:
            raise RuntimeError("Failed to load model with any configuration")
        
        model.eval()
        
        # Set context window
        self.max_context = getattr(model.config, "max_position_embeddings", None) or getattr(self.tokenizer, "model_max_length", 4096)
        if self.max_context is None or self.max_context == int(1e30):
            self.max_context = 4096
        
        # Set token IDs
        self.pad_token_id = self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id
        self.eos_token_id = self.tokenizer.eos_token_id
        
        logger.info(f"Model loaded successfully with {quantization_config} configuration")
        return model
    
    def _build_prompt(self, data: Dict[str, Any]) -> str:
        """Build prompt using chat template or direct input"""
        # Accept either "messages" (chat) or "inputs"/"prompt" (single-turn)
        if "messages" in data and isinstance(data["messages"], list):
            return self.tokenizer.apply_chat_template(
                data["messages"],
                tokenize=False,
                add_generation_prompt=True
            )
        
        user_text = data.get("inputs") or data.get("prompt") or ""
        if isinstance(user_text, str):
            messages = [{"role": "user", "content": user_text}]
            return self.tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=True
            )
        
        return str(user_text)
    
    def _prepare_inputs(self, prompt: str, max_new_tokens: int, params: Dict[str, Any]) -> Dict[str, torch.Tensor]:
        """Prepare inputs with proper tokenization"""
        # Keep room for generation
        max_input_tokens = int(params.get("max_input_tokens", max(self.max_context - max_new_tokens - 8, 256)))
        
        model_inputs = self.tokenizer(
            prompt,
            return_tensors="pt",
            truncation=True,
            max_length=max_input_tokens,
        )
        
        if torch.cuda.is_available():
            model_inputs = {k: v.to(self.model.device) for k, v in model_inputs.items()}
        
        return model_inputs
    
    def _stopping(self, params: Dict[str, Any]):
        """Create stopping criteria"""
        from transformers import StoppingCriteria, StoppingCriteriaList
        
        class StopOnSequences(StoppingCriteria):
            def __init__(self, stop_sequences: List[List[int]]):
                super().__init__()
                self.stop_sequences = [torch.tensor(x, dtype=torch.long) for x in stop_sequences if len(x) > 0]
            
            def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
                if input_ids.shape[0] == 0 or not self.stop_sequences:
                    return False
                generated = input_ids[0]
                for seq in self.stop_sequences:
                    if generated.shape[0] >= seq.shape[0] and torch.equal(generated[-seq.shape[0]:], seq.to(generated.device)):
                        return True
                return False
        
        stop = params.get("stop", [])
        if isinstance(stop, str):
            stop = [stop]
        if not isinstance(stop, list):
            stop = []
        
        stop_ids = [self.tokenizer.encode(s, add_special_tokens=False) for s in stop]
        criteria = []
        if stop_ids:
            criteria.append(StopOnSequences(stop_ids))
        
        return StoppingCriteriaList(criteria)
    
    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """Handle inference requests with proper error handling"""
        try:
            params = data.get("parameters", {}) or {}
            
            # Set seed if provided
            seed = params.get("seed")
            if seed is not None:
                try:
                    torch.manual_seed(int(seed))
                except (ValueError, TypeError):
                    pass
            
            # Generation parameters
            max_new_tokens = int(params.get("max_new_tokens", 512))
            temperature = float(params.get("temperature", 0.2))
            top_p = float(params.get("top_p", 0.9))
            top_k = int(params.get("top_k", 50))
            repetition_penalty = float(params.get("repetition_penalty", 1.05))
            num_beams = int(params.get("num_beams", 1))
            do_sample = bool(params.get("do_sample", temperature > 0 and num_beams == 1))
            
            # Build prompt
            prompt = self._build_prompt(data)
            model_inputs = self._prepare_inputs(prompt, max_new_tokens, params)
            input_length = model_inputs["input_ids"].shape[-1]
            
            # Generation kwargs
            gen_kwargs = dict(
                max_new_tokens=max_new_tokens,
                do_sample=do_sample,
                temperature=max(0.0, temperature),
                top_p=top_p,
                top_k=top_k,
                repetition_penalty=repetition_penalty,
                num_beams=num_beams,
                eos_token_id=self.eos_token_id,
                pad_token_id=self.pad_token_id,
                stopping_criteria=self._stopping(params),
            )
            
            # Generate
            with torch.no_grad():
                output_ids = self.model.generate(**model_inputs, **gen_kwargs)
            
            # Slice off the prompt
            gen_ids = output_ids[0][input_length:]
            text = self.tokenizer.decode(gen_ids, skip_special_tokens=True)
            
            # Apply text-side stop strings if provided
            stop = params.get("stop", [])
            if isinstance(stop, str):
                stop = [stop]
            for s in stop or []:
                idx = text.find(s)
                if idx != -1:
                    text = text[:idx]
                    break
            
            # Token accounting
            prompt_tokens = int(input_length)
            completion_tokens = int(gen_ids.shape[-1])
            total_tokens = prompt_tokens + completion_tokens
            
            return {
                "generated_text": text,
                "input_tokens": prompt_tokens,
                "generated_tokens": completion_tokens,
                "total_tokens": total_tokens,
                "params": {
                    "max_new_tokens": max_new_tokens,
                    "temperature": temperature,
                    "top_p": top_p,
                    "top_k": top_k,
                    "repetition_penalty": repetition_penalty,
                    "num_beams": num_beams,
                    "do_sample": do_sample,
                },
            }
            
        except Exception as e:
            logger.error(f"Generation error: {str(e)}")
            return {
                "error": f"Generation failed: {str(e)}", 
                "generated_text": "",
                "input_tokens": 0,
                "generated_tokens": 0,
                "total_tokens": 0
            }