batch `loglikelihood_rolling` across requests (#2559)

* batch all rolling token windows

* nit

* copy to vllm

* fix max_length for `get_rolling_token_windows`

* bugfix

* bugfix

* add type hints

lm_eval/models/huggingface.py

@@ -905,8 +905,6 @@ class HFLM(TemplateLM):
     def loglikelihood_rolling(
         self, requests: List[Instance], disable_tqdm: bool = False
     ) -> List[float]:
-        loglikelihoods = []
-
         adaptive_batch_size = None
         if self.batch_size == "auto":
             # using rolling window with maximum context
@@ -915,10 +913,17 @@ class HFLM(TemplateLM):
             print(f"Determined Largest batch size: {batch_size}")
             adaptive_batch_size = batch_size
 
-        for (string,) in tqdm(
-            [req.args for req in requests], disable=(disable_tqdm or (self.rank != 0))
+        # First, collect all windows from all requests
+        all_windows = []  # List of (request_idx, window) tuples
+        request_window_counts = []  # Track number of windows per request
+
+        for req_idx, (string,) in enumerate(
+            tqdm(
+                [req.args for req in requests],
+                disable=(disable_tqdm or (self.rank != 0)),
+            )
         ):
-            rolling_token_windows = list(
+            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
                 map(
                     utils.make_disjoint_window,
                     utils.get_rolling_token_windows(
@@ -931,37 +936,55 @@ class HFLM(TemplateLM):
             )
 
             # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
-            rolling_token_windows = [(None,) + x for x in rolling_token_windows]
+            windows = [(None,) + x for x in rolling_token_windows]
+
+            # Store windows with their request index
+            all_windows.extend((req_idx, window) for window in windows)
+            request_window_counts.append(len(windows))
 
+        # Handle distributed case padding
         pad_amnt = 0
         if self.world_size > 1:
-                # We pad out the external document-level iterator so the inner iterator doesn't hang
-                mytensor = torch.tensor(len(rolling_token_windows), device=self.device)
-                gathered = (
-                    self.accelerator.gather(mytensor).cpu().detach().numpy().tolist()
-                )
-
+            mytensor = torch.tensor(len(all_windows), device=self.device)
+            gathered = self.accelerator.gather(mytensor).cpu().detach().numpy().tolist()
             pad_amnt = max(gathered) - gathered[self.rank]
             if pad_amnt > 0:
-                    rolling_token_windows += pad_amnt * [rolling_token_windows[0]]
+                all_windows += pad_amnt * [all_windows[0]]
+
+        all_nlls = []
+        batch_size = adaptive_batch_size or self.batch_size
+        for i in range(0, len(all_windows), batch_size):
+            batch = all_windows[i : i + batch_size]
+            # Extract just the windows for processing, keeping track of request indices
+            batch_indices, batch_windows = zip(*batch)
 
-            string_nll = self._loglikelihood_tokens(
-                requests=rolling_token_windows,
-                disable_tqdm=True,
-                override_bs=adaptive_batch_size,
+            batch_nlls = self._loglikelihood_tokens(
+                requests=batch_windows,
+                disable_tqdm=False,
+                override_bs=len(batch_windows),
             )
+            # Store results with their request indices
+            all_nlls.extend(zip(batch_indices, batch_nlls))
 
+        # Remove padding if necessary
         if (self.world_size > 1) and (pad_amnt > 0):
-                string_nll = [x[0] for x in string_nll[:-pad_amnt]]
-            else:
-                # discard is_greedy
-                string_nll = [x[0] for x in string_nll]
-
-            string_nll = sum(string_nll)
-            loglikelihoods.append(string_nll)
+            all_nlls = all_nlls[:-pad_amnt]
 
-            # cache this loglikelihood_rolling request
-            self.cache_hook.add_partial("loglikelihood_rolling", (string,), string_nll)
+        # Reconstruct per-request loglikelihoods
+        loglikelihoods = []
+        current_idx = 0
+        for window_count in request_window_counts:
+            # Get all nlls for this request
+            request_nlls = all_nlls[current_idx : current_idx + window_count]
+            # Sum up the nlls for this request (discarding is_greedy)
+            request_total = sum(nll[0] for _, nll in request_nlls)
+            loglikelihoods.append(request_total)
+            current_idx += window_count
+
+            string = requests[len(loglikelihoods) - 1].args[0]
+            self.cache_hook.add_partial(
+                "loglikelihood_rolling", (string,), request_total
+            )
 
         return loglikelihoods
 

lm_eval/models/vllm_causallms.py

@@ -102,7 +102,7 @@ class VLLM(TemplateLM):
         self.batch_size = (
             "auto"
             if isinstance(batch_size, str) and "auto" in batch_size
-            else batch_size
+            else int(batch_size)
         )
         if self.data_parallel_size <= 1:
             self.model = LLM(**self.model_args)
@@ -281,10 +281,21 @@ class VLLM(TemplateLM):
     def loglikelihood_rolling(
         self, requests: List[Instance], disable_tqdm: bool = False
     ) -> List[float]:
-        loglikelihoods = []
+        adaptive_batch_size = None
+        if self.batch_size == "auto":
+            adaptive_batch_size = len(requests)
+
+        # First, collect all windows from all requests
+        all_windows = []  # List of (request_idx, window) tuples
+        request_window_counts = []  # Track number of windows per request
 
-        for (string,) in tqdm([req.args for req in requests], disable=disable_tqdm):
-            rolling_token_windows = list(
+        for req_idx, (string,) in enumerate(
+            tqdm(
+                [req.args for req in requests],
+                disable=(disable_tqdm or (self.rank != 0)),
+            )
+        ):
+            rolling_token_windows: List[Tuple[List[int], List[int]]] = list(
                 map(
                     make_disjoint_window,
                     get_rolling_token_windows(
@@ -297,20 +308,42 @@ class VLLM(TemplateLM):
                 )
             )
 
-            rolling_token_windows = [(None,) + x for x in rolling_token_windows]
+            # TODO: Right now, we pass single EOT token to the Encoder and the full context to the decoder, in seq2seq case
+            windows = [(None,) + x for x in rolling_token_windows]
 
-            string_nll = self._loglikelihood_tokens(
-                rolling_token_windows,
-            )
+            # Store windows with their request index
+            all_windows.extend((req_idx, window) for window in windows)
+            request_window_counts.append(len(windows))
 
-            # discard is_greedy
-            string_nll = [x[0] for x in string_nll]
+        all_nlls = []
+        batch_size = adaptive_batch_size or int(self.batch_size)
+        for i in range(0, len(all_windows), batch_size):
+            batch = all_windows[i : i + batch_size]
+            # Extract just the windows for processing, keeping track of request indices
+            batch_indices, batch_windows = zip(*batch)
 
-            string_nll = sum(string_nll)
-            loglikelihoods.append(string_nll)
+            batch_nlls = self._loglikelihood_tokens(
+                requests=batch_windows,
+                disable_tqdm=False,
+            )
+            # Store results with their request indices
+            all_nlls.extend(zip(batch_indices, batch_nlls))
 
-            # cache this loglikelihood_rolling request
-            self.cache_hook.add_partial("loglikelihood_rolling", (string,), string_nll)
+        # Reconstruct per-request loglikelihoods
+        loglikelihoods = []
+        current_idx = 0
+        for window_count in request_window_counts:
+            # Get all nlls for this request
+            request_nlls = all_nlls[current_idx : current_idx + window_count]
+            # Sum up the nlls for this request (discarding is_greedy)
+            request_total = sum(nll[0] for _, nll in request_nlls)
+            loglikelihoods.append(request_total)
+            current_idx += window_count
+
+            string = requests[len(loglikelihoods) - 1].args[0]
+            self.cache_hook.add_partial(
+                "loglikelihood_rolling", (string,), request_total
+            )
 
         return loglikelihoods
 

lm_eval/utils.py

@@ -10,7 +10,7 @@ import os
 import re
 from dataclasses import asdict, is_dataclass
 from itertools import islice
-from typing import Any, Callable, List
+from typing import Any, Callable, Generator, List, Tuple
 
 import numpy as np
 import yaml
@@ -201,7 +201,9 @@ def get_sample_results_filenames(filenames: List[str]) -> List[str]:
     return [f for f in filenames if "/samples_" in f and ".json" in f]
 
 
-def get_rolling_token_windows(token_list, prefix_token, max_seq_len, context_len):
+def get_rolling_token_windows(
+    token_list: List[int], prefix_token: int, max_seq_len: int, context_len: int
+) -> Generator[Tuple[List[int], List[int]], None, None]:
     """
     - context_len allows for a rolling window context, allowing each prediction window to potentially
       condition on some context
@@ -228,7 +230,7 @@ def get_rolling_token_windows(token_list, prefix_token, max_seq_len, context_len
 
     # Special handling for first window: predict all tokens
     first_seq_len = min(max_seq_len, len(token_list))
-    yield ([prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len])
+    yield [prefix_token] + token_list[: first_seq_len - 1], token_list[:first_seq_len]
     predicted += first_seq_len
 
     while predicted < len(token_list):
@@ -242,7 +244,9 @@ def get_rolling_token_windows(token_list, prefix_token, max_seq_len, context_len
         predicted += window_pred_len
 
 
-def make_disjoint_window(pair):
+def make_disjoint_window(
+    pair: Tuple[List[int], List[int]],
+) -> Tuple[List[int], List[int]]:
     """Takes output from get_rolling_token_windows and makes the context not overlap with the continuation"""
     a, b = pair
     return a[: len(a) - (len(b) - 1)], b