Support streaming on internal API

languagemodels/inference.py

@@ -112,6 +112,18 @@ def chat_oa(engine, prompt, max_tokens=200, temperature=0):
         raise InferenceException(f"OpenAI error: {resp}")
 
 
+def stream_results(results, tokenizer):
+    """Map a token iterator to a substring iterator"""
+    tokens = []
+    last_len = 0
+
+    for result in results:
+        tokens.append(result.token_id)
+        text = tokenizer.decode(tokens)
+        yield text[last_len:]
+        last_len = len(text)
+
+
 def generate(
     instructions: List[str],
     max_tokens: int = 200,
@@ -121,6 +133,7 @@ def generate(
     prefix: str = "",
     suppress: List[str] = [],
     model: str = "instruct",
+    stream: bool = False,
 ):
     """Generates completions for a prompt
 
@@ -129,6 +142,9 @@ def generate(
 
     >>> generate(["What is the capital of France?"])
     ['...Paris...']
+
+    >>> list(generate(["What is the capital of France?"], stream=True))
+    ['...Paris...']
     """
     if os.environ.get("LANGUAGEMODELS_TS_KEY") or os.environ.get(
         "LANGUAGEMODELS_TS_SERVER"
@@ -157,31 +173,56 @@ def generate(
     outputs_ids = []
     if hasattr(model, "translate_batch"):
         prefix = tokenizer.encode(prefix, add_special_tokens=False).tokens
-        results = model.translate_batch(
-            prompts_tok,
-            target_prefix=[prefix] * len(prompts),
-            repetition_penalty=repetition_penalty,
-            max_decoding_length=max_tokens,
-            sampling_temperature=temperature,
-            sampling_topk=topk,
-            suppress_sequences=suppress,
-            beam_size=1,
-        )
-        outputs_tokens = [r.hypotheses[0] for r in results]
-        for output in outputs_tokens:
-            outputs_ids.append([tokenizer.token_to_id(t) for t in output])
+        if stream:
+            results = model.generate_tokens(
+                prompts_tok[0],
+                target_prefix=prefix,
+                repetition_penalty=repetition_penalty,
+                max_decoding_length=max_tokens,
+                sampling_temperature=temperature,
+                sampling_topk=topk,
+                suppress_sequences=suppress,
+            )
+
+            return stream_results(results, tokenizer)
+        else:
+            results = model.translate_batch(
+                prompts_tok,
+                target_prefix=[prefix] * len(prompts),
+                repetition_penalty=repetition_penalty,
+                max_decoding_length=max_tokens,
+                sampling_temperature=temperature,
+                sampling_topk=topk,
+                suppress_sequences=suppress,
+                beam_size=1,
+            )
+            outputs_tokens = [r.hypotheses[0] for r in results]
+            for output in outputs_tokens:
+                outputs_ids.append([tokenizer.token_to_id(t) for t in output])
     else:
-        results = model.generate_batch(
-            prompts_tok,
-            repetition_penalty=repetition_penalty,
-            max_length=max_tokens,
-            sampling_temperature=temperature,
-            sampling_topk=topk,
-            suppress_sequences=suppress,
-            beam_size=1,
-            include_prompt_in_result=False,
-        )
-        outputs_ids = [r.sequences_ids[0] for r in results]
+        if stream:
+            results = model.generate_tokens(
+                prompts_tok,
+                repetition_penalty=repetition_penalty,
+                max_length=max_tokens,
+                sampling_temperature=temperature,
+                sampling_topk=topk,
+                suppress_sequences=suppress,
+            )
+
+            return stream_results(results, tokenizer)
+        else:
+            results = model.generate_batch(
+                prompts_tok,
+                repetition_penalty=repetition_penalty,
+                max_length=max_tokens,
+                sampling_temperature=temperature,
+                sampling_topk=topk,
+                suppress_sequences=suppress,
+                beam_size=1,
+                include_prompt_in_result=False,
+            )
+            outputs_ids = [r.sequences_ids[0] for r in results]
 
     model_info["requests"] = model_info.get("requests", 0) + len(prompts)