Merge pull request #1 from fjfricke/inline_probabilities

add first version where you can get logprobs from tokens in argmax when using where in set

Author: fjfricke

src/lmql/runtime/dclib/dclib_array.py

@@ -12,6 +12,7 @@ class Continuation:
     token: Any
     logprob: Any
     user_data: Any
+    distribution_logprobs: Any = None
 class criterion: 
     def __and__(self, other):
         return logical_and(self, other)
@@ -325,8 +326,9 @@ def op_extend(p1, p2):
                 tokens = continuation.token.reshape(-1)
                 logprobs = continuation.logprob.reshape(-1)
                 user_data = continuation.user_data or [None] * len(tokens)
-                for t,s,u in zip(tokens, logprobs, user_data):
-                    extended_seqs.append(sq.extend(Continuation(t, s, u)))
+                distribution_logprobs = continuation.distribution_logprobs or [None] * len(tokens)
+                for t,s,u,d in zip(tokens, logprobs, user_data, distribution_logprobs):
+                    extended_seqs.append(sq.extend(Continuation(t, s, u, d)))
             return extended_seqs
 
         return DataArray(apply_componentwise(op_extend, self.sequences, other.sequences, "extend", allow_mismatch_keys=False), dims=self.shape)

src/lmql/runtime/dclib/dclib_seq.py

@@ -82,7 +82,7 @@ async def json(self, diff: bool = False):
         }
 
 class DecoderSequence:
-    def __init__(self, input_ids_or_str, logprobs=None, deterministic=None, stop_phrase=None, predecessor=None, user_data=None, sticky_user_data_keys=None, epsilon_node=False, internal=False):
+    def __init__(self, input_ids_or_str, logprobs=None, deterministic=None, stop_phrase=None, predecessor=None, user_data=None, sticky_user_data_keys=None, epsilon_node=False, internal=False, distribution_logprobs=None):
         if logprobs is not None:
             if not all([p > get_truncation_threshold() for p in logprobs]):
                 warnings.warn("logprobs contain values below the current logprob truncation threshold {t}, which may cause unexpected behavior. Consider increasing the truncation threshold via lmql.model(..., truncation_threshold=...).".format(t=get_truncation_threshold()))
@@ -141,6 +141,10 @@ def __init__(self, input_ids_or_str, logprobs=None, deterministic=None, stop_phr
 
         # indicates to dc.rewrite whether this sequence can be rewritten
         self.needs_rewrite = True
+        if not distribution_logprobs:
+            self.distribution_logprobs = [None] * len(self.logprobs)
+        else:
+            self.distribution_logprobs = distribution_logprobs
 
     def __hash__(self) -> int:
         return hash(self.id)
@@ -371,7 +375,8 @@ def extend(self, continuation, internal=False):
             predecessor=self,
             user_data=self.extend_user_data(continuation),
             sticky_user_data_keys=self.sticky_user_data_keys,
-            internal=internal
+            internal=internal,
+            distribution_logprobs=self.distribution_logprobs + [continuation.distribution_logprobs]
         )
 
     def detect_stop_phrase(self, continuation):
@@ -436,14 +441,18 @@ def make_successors(self, next_tokens, next_token_scores, logits, user_data=None
         tokens = [t for t, s in zip(next_tokens, next_token_scores) if s > get_truncation_threshold()]
         scores = [s for s in next_token_scores if s > get_truncation_threshold()]
 
+        distribution_logprobs = [{k: v for k, v in logits.probs.items() if type(k) == str}]
+        if len(distribution_logprobs[0]) < 1:
+            distribution_logprobs = None
+
         if len(tokens) == 0:
             print("WARNING: all continuation token fall below the current logprob truncation threshold {t}. This is likely due to a too low truncation threshold. Please increase the truncation threshold via lmql.model(..., truncation_threshold=...).".format(t=get_truncation_threshold()))
             tokens = [t for t, s in zip(next_tokens, next_token_scores)][:1]
             scores = [s for s in next_token_scores][:1]
         next_tokens = np.stack(tokens, axis=0)
         next_token_scores = np.stack(scores, axis=0)
 
-        return Continuation(next_tokens, next_token_scores, user_data)
+        return Continuation(next_tokens, next_token_scores, user_data, distribution_logprobs)
 # global counter for all sequences created in this process for identification purposes
 DecoderSequence.seq_ctr = 0
 DecoderSequence.graph = None

src/lmql/runtime/interpreter.py

@@ -102,6 +102,8 @@ class PromptState(NamedTuple):
     where: Optional[Any]
     tail: Optional[str]
 
+    distribution_logprobs: Optional[Dict[str, float]] = {}
+
     def __str__(self):
         return f"<PromptState '{self.variable}' '{[self.prompt]}'>"
 
@@ -546,15 +548,20 @@ async def where_step_for_sequence(self, s: dc.DecoderSequence, needs_masking, se
             # update hint for max_tokens to generate for current var
             max_tokens_hint = ops.most_restrictive_hint([sub_max_token_hints, max_tokens_hint])
 
+            if len(s.distribution_logprobs) > variable_offset:
+                scores = s.distribution_logprobs[variable_offset]
+            else:
+                scores = None
+
             # current context
             program_state: ProgramState = state.program_state.copy()
-            program_state.set(variable, text, scores=(), diff=diff_text, montonicity="inc", tokens=text_tokens)
+            program_state.set(variable, text, scores=scores, diff=diff_text, montonicity="inc", tokens=text_tokens)
             program_state.subinterpreter_results = subvalid
             program_state.prompt = state.prompt
 
             # follow context
             follow_program_state: ProgramState = state.program_state.copy()
-            follow_program_state.set(variable, text + str(ops.NextToken), scores=(), diff=diff_text, montonicity="inc", tokens=text_tokens)
+            follow_program_state.set(variable, text + str(ops.NextToken), scores=scores, diff=diff_text, montonicity="inc", tokens=text_tokens)
             follow_program_state.subinterpreter_results = subfollow
             follow_program_state.prompt = state.prompt
 
@@ -614,6 +621,7 @@ async def where_step_for_sequence(self, s: dc.DecoderSequence, needs_masking, se
                         program_state=program_state,
                         stopping_phrases=stopping_phrases,
                         where=await self.where_graph_with_trace(where, trace, follow_trace),
+                        distribution_logprobs=scores,
         )
 
         # extract hint of maximum number of tokens to generate for 'variable' from 
@@ -761,7 +769,7 @@ async def rewrite_for_sequence(self, seq: dc.DecoderSequence, needs_rewrite, ass
 
             variable_value = text
             # set raw variable value
-            program_state.set(variable, variable_value, scores=(), diff=text_diff, montonicity="fin", tokens=text_tokens)
+            program_state.set(variable, variable_value, scores=state.distribution_logprobs, diff=text_diff, montonicity="fin", tokens=text_tokens)
 
             where = state.full_where_condition(self)
 
@@ -1022,6 +1030,9 @@ async def debug_out(decoder_step):
                 if _DCLibDebugPrinter.printer.records_graph:
                     dc.set_record_graph()
                     self.decoder_graph = dc.DecoderSequence.graph
+        if self.model.adapter.decoder_args.get("decoder_graph", False):
+            dc.set_record_graph()
+            self.decoder_graph = dc.DecoderSequence.graph
 
         # get decoder function
         mode = decoder_args["decoder"].lower()

src/lmql/runtime/openai_integration.py

@@ -557,7 +557,7 @@ async def op_sample(seqs):
                 sampling_modes = [f"sample-{temperature}-sample-id-{random.randint(0, 2**32-1)}" for _ in range(len(seqs))]
                 edge_type_populated_user_data = [{"dc-edge-type": sm} for sm in sampling_modes]
 
-            completions: List[CompletionResult] = await self.completion_buffer(seqs, logprobs=num_samples, sampling_modes=sampling_modes, **kwargs)
+            completions: List[CompletionResult] = await self.completion_buffer(seqs, logprobs=5, sampling_modes=sampling_modes, **kwargs)
 
             next_token_ids = []
             next_token_scores = []

test.py

@@ -0,0 +1,48 @@
+import sys
+sys.path.append('/Users/felix/Programming/lmql/src')
+import lmql
+import asyncio
+
+#add replicate api key to env
+import os
+os.environ['REPLICATE_API_TOKEN'] = 'r8_aOlrg82Wfg30Rx4L4mv9wI2npPfBQGO0Pvci4'
+
+# def test_decorator(variable_value, prompt_value, context):
+#     return variable_value, prompt_value
+
+async def main():
+
+    test = lmql.model(
+        "openai/gpt-3.5-turbo-instruct"
+        # "meta-llama/Llama-2-13b-chat-hf",
+        # endpoint="replicate:deployment/ml-delphai/llama2-13b-chat-lmtp",
+        # endpoint="replicate:charles-dyfis-net/llama-2-7b-chat-hf--lmtp-8bit",
+        # tokenizer="AyyYOO/Luna-AI-Llama2-Uncensored-FP16-sharded",
+    )
+    pass
+
+    answer = await lmql.run(
+        """
+        import math
+        def get_probs(variable_value, prompt_value, context):
+            breakpoint()
+            logprob_scores = list(context.variable_scores.items())[-1][1]
+            scores = dict()
+            for key, value in logprob_scores.items():
+                if value > -5:
+                    scores[key] = math.exp(value)
+            return scores
+        argmax(verbose=True)
+        \"How much you like monkeys between 0 and 2?[@get_probs MONKEY]\" where MONKEY in set ([\"0\", \"1\", \"2\"])
+        \"How much you like birds between 0 and 2?[@get_probs BIRD]\" where BIRD in set ([\"0\", \"1\", \"2\"])
+        return (MONKEY, BIRD)
+        """,
+        max_len=4000,
+        model=test,
+        # decoder_graph=True,
+    )
+
+    print(answer)
+
+if __name__ == "__main__":
+    asyncio.run(main())