Skip to content

model

optimus_dl.modules.eval.model

LLMBaselinesModel

Bases: LM

LLM Baselines evaluation model for llm_harness.

This class implements the lm_eval interface using a pre-loaded model and tokenizer. All checkpoint loading logic is handled by the EvalRecipe.

Source code in optimus_dl/modules/eval/model.py 
 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
class LLMBaselinesModel(LM):
    """LLM Baselines evaluation model for llm_harness.

    This class implements the lm_eval interface using a pre-loaded model and tokenizer.
    All checkpoint loading logic is handled by the EvalRecipe.
    """

    def __init__(
        self,
        model: BaseModel,
        tokenizer,
        tokenizer_config,
        device: str | torch.device,
    ):
        """Initialize the model with pre-loaded components.

        Args:
            model: Pre-loaded BaseModel instance
            tokenizer: Pre-loaded tokenizer instance
            tokenizer_config: Tokenizer configuration for type detection
            device: Device the model is running on
        """
        super().__init__()

        self.model = model
        self.tokenizer = tokenizer
        self.tokenizer_config = tokenizer_config
        self.device = device

        logger.info(f"LLMBaselinesModel initialized on {self.device}")

    def _tokenize(self, text: str) -> torch.Tensor:
        encoded = self.tokenizer.encode(text)
        return torch.as_tensor(encoded, dtype=torch.long)

    def _decode_tokens(self, tokens: list[int]) -> str:
        """Decode tokens to text using the model's tokenizer."""
        return self.tokenizer.decode(tokens)

    def _compute_logprobs(self, input_ids: torch.Tensor) -> torch.Tensor:
        """Compute log probabilities for input tokens."""
        input_ids = input_ids.to(self.device)

        with torch.no_grad():
            if input_ids.dim() == 1:
                input_ids = input_ids.unsqueeze(0)

            outputs = self.model(input_ids)
            logits = outputs["logits"]  # Shape: [batch, seq_len, vocab_size]

            # Convert to log probabilities
            log_probs = F.log_softmax(logits, dim=-1)

        return log_probs

    def loglikelihood(self, requests: list[Instance]) -> list[tuple[float, bool]]:
        """Compute log-likelihood of generating continuations from contexts.

        Args:
            requests: List of Instance objects with (context, continuation) pairs

        Returns:
            List of (logprob, is_greedy) tuples
        """
        results = []

        for request in tqdm(requests, desc="Computing loglikelihood", leave=False):
            args = request.args
            if not isinstance(args, tuple) or len(args) != 2:
                raise ValueError(f"Expected 2 arguments for loglikelihood, got {args}")
            context, continuation = args

            # Tokenize context and continuation
            context_tokens = self._tokenize(context)
            continuation_tokens = self._tokenize(continuation)

            # Combine for full sequence
            full_tokens = torch.cat([context_tokens, continuation_tokens])

            # Get log probabilities
            log_probs = self._compute_logprobs(full_tokens)

            # Calculate log probability of continuation
            context_len = len(context_tokens)
            continuation_len = len(continuation_tokens)

            if continuation_len == 0:
                logprob = 0.0
                is_greedy = True
            else:
                # Get logprobs for continuation tokens
                relevant_logprobs = log_probs[
                    0, context_len - 1 : context_len + continuation_len - 1
                ]
                token_logprobs = relevant_logprobs[
                    torch.arange(continuation_len), continuation_tokens
                ]
                logprob = token_logprobs.sum().item()

                # Check if this would be the greedy choice
                greedy_tokens = log_probs[
                    0, context_len - 1 : context_len + continuation_len - 1
                ].argmax(dim=-1)
                is_greedy = torch.equal(
                    greedy_tokens, continuation_tokens.to(greedy_tokens.device)
                )

            results.append((logprob, is_greedy))

        return results

    def loglikelihood_rolling(self, requests: list[Instance]) -> list[float]:
        """Compute rolling log-likelihood for perplexity calculation.

        Args:
            requests: List of Instance objects with (text,) tuples

        Returns:
            List of log probabilities
        """
        results = []

        for request in requests:
            args = request.args
            if not isinstance(args, tuple) or len(args) != 1:
                raise ValueError(
                    f"Expected 1 argument for loglikelihood_rolling, got {args}"
                )
            text = args[0]

            # Tokenize text
            tokens = self._tokenize(text)

            if len(tokens) <= 1:
                results.append(0.0)
                continue

            # Get log probabilities
            log_probs = self._compute_logprobs(tokens)

            # Calculate total log probability (excluding first token)
            token_indices = tokens[1:]  # Target tokens
            context_logprobs = log_probs[0, :-1]  # Logprobs from context positions

            selected_logprobs = context_logprobs[
                torch.arange(len(token_indices)), token_indices
            ]
            total_logprob = selected_logprobs.sum().item()

            results.append(total_logprob)

        return results

    def generate_until(self, requests: list[Instance]) -> list[str]:
        """Generate text until stopping criteria are met.

        Args:
            requests: List of Instance objects with (context, gen_kwargs) pairs

        Returns:
            List of generated text continuations
        """
        results = []

        for request in requests:
            args = request.args
            if not isinstance(args, tuple) or len(args) != 2:
                raise ValueError(f"Expected 2 arguments for generate_until, got {args}")
            context, gen_kwargs = args

            # Parse generation arguments
            max_gen_toks = gen_kwargs.get("max_gen_toks", 256)
            until = gen_kwargs.get("until", [])
            temperature = gen_kwargs.get("temperature", 0.0)  # Greedy by default

            # Tokenize context
            context_tokens = self._tokenize(context)
            input_ids = context_tokens.unsqueeze(0).to(self.device)

            # Generate tokens
            generated_tokens = []

            with torch.no_grad():
                for _ in range(max_gen_toks):
                    # Get next token logits
                    outputs = self.model(input_ids)
                    logits = outputs["logits"][:, -1, :]  # Last position

                    # Sample next token
                    if temperature == 0.0:
                        next_token = logits.argmax(dim=-1)
                    else:
                        probs = F.softmax(logits / temperature, dim=-1)
                        next_token = torch.multinomial(probs, 1).squeeze(-1)

                    generated_tokens.append(next_token.item())

                    # Update input for next iteration
                    input_ids = torch.cat(
                        [input_ids, next_token.unsqueeze(0).unsqueeze(0)], dim=1
                    )

                    # Check stopping criteria
                    current_text = self._decode_tokens(generated_tokens)
                    if any(stop_seq in current_text for stop_seq in until):
                        # Truncate at first stopping sequence
                        for stop_seq in until:
                            if stop_seq in current_text:
                                current_text = current_text.split(stop_seq)[0]
                                break
                        break

            generated_text = (
                self._decode_tokens(generated_tokens) if generated_tokens else ""
            )

            # Remove stopping sequences from the end
            for stop_seq in until:
                if generated_text.endswith(stop_seq):
                    generated_text = generated_text[: -len(stop_seq)]

            results.append(generated_text)

        return results

    @property
    def tokenizer_name(self) -> str:
        """Return tokenizer name for caching."""
        return f"{self.tokenizer_config.type}:{self.tokenizer_config.name}"

tokenizer_name property

Return tokenizer name for caching.

__init__(model, tokenizer, tokenizer_config, device)

Initialize the model with pre-loaded components.

Parameters:

Name Type Description Default
model BaseModel

Pre-loaded BaseModel instance

 required
tokenizer

Pre-loaded tokenizer instance

 required
tokenizer_config

Tokenizer configuration for type detection

 required
device str | device

Device the model is running on

 required
Source code in optimus_dl/modules/eval/model.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
def __init__(
    self,
    model: BaseModel,
    tokenizer,
    tokenizer_config,
    device: str | torch.device,
):
    """Initialize the model with pre-loaded components.

    Args:
        model: Pre-loaded BaseModel instance
        tokenizer: Pre-loaded tokenizer instance
        tokenizer_config: Tokenizer configuration for type detection
        device: Device the model is running on
    """
    super().__init__()

    self.model = model
    self.tokenizer = tokenizer
    self.tokenizer_config = tokenizer_config
    self.device = device

    logger.info(f"LLMBaselinesModel initialized on {self.device}")

generate_until(requests)

Generate text until stopping criteria are met.

Parameters:

Name Type Description Default
requests list[Instance]

List of Instance objects with (context, gen_kwargs) pairs

 required

Returns:

Type Description
list[str]

List of generated text continuations
Source code in optimus_dl/modules/eval/model.py 
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
def generate_until(self, requests: list[Instance]) -> list[str]:
    """Generate text until stopping criteria are met.

    Args:
        requests: List of Instance objects with (context, gen_kwargs) pairs

    Returns:
        List of generated text continuations
    """
    results = []

    for request in requests:
        args = request.args
        if not isinstance(args, tuple) or len(args) != 2:
            raise ValueError(f"Expected 2 arguments for generate_until, got {args}")
        context, gen_kwargs = args

        # Parse generation arguments
        max_gen_toks = gen_kwargs.get("max_gen_toks", 256)
        until = gen_kwargs.get("until", [])
        temperature = gen_kwargs.get("temperature", 0.0)  # Greedy by default

        # Tokenize context
        context_tokens = self._tokenize(context)
        input_ids = context_tokens.unsqueeze(0).to(self.device)

        # Generate tokens
        generated_tokens = []

        with torch.no_grad():
            for _ in range(max_gen_toks):
                # Get next token logits
                outputs = self.model(input_ids)
                logits = outputs["logits"][:, -1, :]  # Last position

                # Sample next token
                if temperature == 0.0:
                    next_token = logits.argmax(dim=-1)
                else:
                    probs = F.softmax(logits / temperature, dim=-1)
                    next_token = torch.multinomial(probs, 1).squeeze(-1)

                generated_tokens.append(next_token.item())

                # Update input for next iteration
                input_ids = torch.cat(
                    [input_ids, next_token.unsqueeze(0).unsqueeze(0)], dim=1
                )

                # Check stopping criteria
                current_text = self._decode_tokens(generated_tokens)
                if any(stop_seq in current_text for stop_seq in until):
                    # Truncate at first stopping sequence
                    for stop_seq in until:
                        if stop_seq in current_text:
                            current_text = current_text.split(stop_seq)[0]
                            break
                    break

        generated_text = (
            self._decode_tokens(generated_tokens) if generated_tokens else ""
        )

        # Remove stopping sequences from the end
        for stop_seq in until:
            if generated_text.endswith(stop_seq):
                generated_text = generated_text[: -len(stop_seq)]

        results.append(generated_text)

    return results

loglikelihood(requests)

Compute log-likelihood of generating continuations from contexts.

Parameters:

Name Type Description Default
requests list[Instance]

List of Instance objects with (context, continuation) pairs

 required

Returns:

Type Description
list[tuple[float, bool]]

List of (logprob, is_greedy) tuples
Source code in optimus_dl/modules/eval/model.py 
 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
def loglikelihood(self, requests: list[Instance]) -> list[tuple[float, bool]]:
    """Compute log-likelihood of generating continuations from contexts.

    Args:
        requests: List of Instance objects with (context, continuation) pairs

    Returns:
        List of (logprob, is_greedy) tuples
    """
    results = []

    for request in tqdm(requests, desc="Computing loglikelihood", leave=False):
        args = request.args
        if not isinstance(args, tuple) or len(args) != 2:
            raise ValueError(f"Expected 2 arguments for loglikelihood, got {args}")
        context, continuation = args

        # Tokenize context and continuation
        context_tokens = self._tokenize(context)
        continuation_tokens = self._tokenize(continuation)

        # Combine for full sequence
        full_tokens = torch.cat([context_tokens, continuation_tokens])

        # Get log probabilities
        log_probs = self._compute_logprobs(full_tokens)

        # Calculate log probability of continuation
        context_len = len(context_tokens)
        continuation_len = len(continuation_tokens)

        if continuation_len == 0:
            logprob = 0.0
            is_greedy = True
        else:
            # Get logprobs for continuation tokens
            relevant_logprobs = log_probs[
                0, context_len - 1 : context_len + continuation_len - 1
            ]
            token_logprobs = relevant_logprobs[
                torch.arange(continuation_len), continuation_tokens
            ]
            logprob = token_logprobs.sum().item()

            # Check if this would be the greedy choice
            greedy_tokens = log_probs[
                0, context_len - 1 : context_len + continuation_len - 1
            ].argmax(dim=-1)
            is_greedy = torch.equal(
                greedy_tokens, continuation_tokens.to(greedy_tokens.device)
            )

        results.append((logprob, is_greedy))

    return results

loglikelihood_rolling(requests)

Compute rolling log-likelihood for perplexity calculation.

Parameters:

Name Type Description Default
requests list[Instance]

List of Instance objects with (text,) tuples

 required

Returns:

Type Description
list[float]

List of log probabilities
Source code in optimus_dl/modules/eval/model.py 
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
def loglikelihood_rolling(self, requests: list[Instance]) -> list[float]:
    """Compute rolling log-likelihood for perplexity calculation.

    Args:
        requests: List of Instance objects with (text,) tuples

    Returns:
        List of log probabilities
    """
    results = []

    for request in requests:
        args = request.args
        if not isinstance(args, tuple) or len(args) != 1:
            raise ValueError(
                f"Expected 1 argument for loglikelihood_rolling, got {args}"
            )
        text = args[0]

        # Tokenize text
        tokens = self._tokenize(text)

        if len(tokens) <= 1:
            results.append(0.0)
            continue

        # Get log probabilities
        log_probs = self._compute_logprobs(tokens)

        # Calculate total log probability (excluding first token)
        token_indices = tokens[1:]  # Target tokens
        context_logprobs = log_probs[0, :-1]  # Logprobs from context positions

        selected_logprobs = context_logprobs[
            torch.arange(len(token_indices)), token_indices
        ]
        total_logprob = selected_logprobs.sum().item()

        results.append(total_logprob)

    return results