Merge pull request #80 from BatsResearch/modularize_static_batching

trigger do_sample automatically based on temperature for huggingface …
BatsResearch · Jul 21, 2024 · 8847315 · 8847315
2 parents 8abca96 + 7fca0f1
commit 8847315
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 11 deletions.
diff --git a/alfred/fm/huggingface.py b/alfred/fm/huggingface.py
@@ -130,7 +130,9 @@ def __init__(
 
         if torch.cuda.is_available():
             n_gpus = torch.cuda.device_count()
-            free_in_GB = sum([int(mem / 1024**3) for mem in torch.cuda.mem_get_info()])
+            free_in_GB = sum(
+                [int(mem / 1024**3) for mem in torch.cuda.mem_get_info()]
+            )
 
             logger.log(
                 logging.INFO, f"Found {n_gpus} GPUs with {free_in_GB}GB free GPU memory"
@@ -398,9 +400,10 @@ def _generate_batch(
                 outputs = self.model.generate(
                     inputs.input_ids.to(self.model.device),
                     max_new_tokens=max_new_tokens,
-                    temperature=temprature,
+                    temperature=temprature if temprature != 0 else None,
                     repetition_penalty=repetition_penalty,
                     return_dict_in_generate=True,
+                    do_sample=temprature != 0,
                 )
             else:
                 outputs = [

diff --git a/alfred/fm/utils.py b/alfred/fm/utils.py
@@ -526,13 +526,15 @@ def _process_batch(batch):
         return batches
 
 
-def static_batch(queries: Query, batch_size: int = 1024) -> List[List[Query]]:
+def static_batch(
+    queries: Union[Query, str], batch_size: int = 512
+) -> List[List[Query]]:
     """
     Static Batching Utility
     Batch queries into fixed size batches
 
     :param queries: A list of queries to be batched
-    :type queries: List[Query]
+    :type queries: Union[Query, str]
     :param batch_sz: The batch size
     :type batch_sz: int
     :return: A list of batches
@@ -548,6 +550,10 @@ def static_batch(queries: Query, batch_size: int = 1024) -> List[List[Query]]:
             _q = query.load()[0]
         elif isinstance(query, RankedQuery):
             _q = query.prompt
+        elif isinstance(query, str):
+            _q = query
+        else:
+            print(f"Unknown query type {type(query)}")
         batch.append(_q)
     if len(batch) > 0:
         batches.append(batch)

diff --git a/docs/alfred/fm/huggingface.md b/docs/alfred/fm/huggingface.md
@@ -47,7 +47,7 @@ class HuggingFaceModel(LocalAccessFoundationModel):
 
 ### HuggingFaceModel()._encode_batch
 
-[Show source in huggingface.py:438](../../../alfred/fm/huggingface.py#L438)
+[Show source in huggingface.py:441](../../../alfred/fm/huggingface.py#L441)
 
 Encode given batch of instances.
 
@@ -71,7 +71,7 @@ def _encode_batch(self, batch_instance, **kwargs) -> List[torch.Tensor]: ...
 
 ### HuggingFaceModel()._generate_batch
 
-[Show source in huggingface.py:348](../../../alfred/fm/huggingface.py#L348)
+[Show source in huggingface.py:350](../../../alfred/fm/huggingface.py#L350)
 
 Generate completions for a batch of prompts using the model.
 
@@ -114,7 +114,7 @@ def _generate_batch(
 
 ### HuggingFaceModel()._get_hidden_states
 
-[Show source in huggingface.py:173](../../../alfred/fm/huggingface.py#L173)
+[Show source in huggingface.py:175](../../../alfred/fm/huggingface.py#L175)
 
 Get the hidden states of the inputs.
 For encoder-decoder models (e.g.) T5, this returns the encoder hidden states.
@@ -140,7 +140,7 @@ def _get_hidden_states(self, inputs, reduction="mean") -> torch.Tensor: ...
 
 ### HuggingFaceModel()._score_batch
 
-[Show source in huggingface.py:212](../../../alfred/fm/huggingface.py#L212)
+[Show source in huggingface.py:214](../../../alfred/fm/huggingface.py#L214)
 
 Score a batch of prompts and candidates using the model.
 
@@ -180,7 +180,7 @@ def _score_batch(
 
 ### HuggingFaceModel().chat
 
-[Show source in huggingface.py:464](../../../alfred/fm/huggingface.py#L464)
+[Show source in huggingface.py:467](../../../alfred/fm/huggingface.py#L467)
 
 Launch an interactive chat session
 

diff --git a/docs/alfred/fm/utils.md b/docs/alfred/fm/utils.md
@@ -380,7 +380,7 @@ Batch queries into fixed size batches
 #### Arguments
 
 - `queries` - A list of queries to be batched
-:type queries: List[Query]
+:type queries: Union[Query, str]
 - `batch_sz` - The batch size
 :type batch_sz: int
 
@@ -392,7 +392,9 @@ Type: *List[List[Query]]*
 #### Signature
 
 ```python
-def static_batch(queries: Query, batch_sz: int = 1024) -> List[List[Query]]: ...
+def static_batch(
+    queries: Union[Query, str], batch_size: int = 512
+) -> List[List[Query]]: ...
 ```