fix: Handle concurrent grammar requests (#1610)

This PR fixes parallel grammar requests, currently grammar states are
not concatenated correctly when a new request is added to the batch and
this results in incorrect generation. This PR updates the `concatenate`
function to correctly include the previous states.

fixes: #1601

integration-tests/models/__snapshots__/test_grammar_llama/test_flash_llama_grammar_load.json

@@ -61,7 +61,7 @@
         },
         {
           "id": 29906,
-          "logprob": -0.2376709,
+          "logprob": -0.33666992,
           "special": false,
           "text": "2"
         },
@@ -180,7 +180,7 @@
         },
         {
           "id": 29906,
-          "logprob": -0.23840332,
+          "logprob": -0.33740234,
           "special": false,
           "text": "2"
         },
@@ -299,7 +299,7 @@
         },
         {
           "id": 29906,
-          "logprob": -0.23840332,
+          "logprob": -0.33740234,
           "special": false,
           "text": "2"
         },
@@ -418,7 +418,7 @@
         },
         {
           "id": 29906,
-          "logprob": -0.23840332,
+          "logprob": -0.33740234,
           "special": false,
           "text": "2"
         },

server/text_generation_server/models/flash_causal_lm.py

@@ -530,6 +530,7 @@ class FlashCausalLMBatch(Batch):
         read_offsets = []
 
         next_token_chooser_parameters = []
+        fsm_grammar_states = []
         stopping_criterias = []
         top_n_tokens = []
 
@@ -578,6 +579,7 @@ class FlashCausalLMBatch(Batch):
             read_offsets.extend(batch.read_offsets)
 
             next_token_chooser_parameters.extend([r.parameters for r in batch.requests])
+            fsm_grammar_states.extend(batch.next_token_chooser.fsm_grammar_states)
             stopping_criterias.extend(batch.stopping_criterias)
 
             top_n_tokens.extend(batch.top_n_tokens)
@@ -593,6 +595,7 @@ class FlashCausalLMBatch(Batch):
             dtype=batches[0].next_token_chooser.dtype,
             device=batches[0].next_token_chooser.device,
             tokenizer=batches[0].next_token_chooser.tokenizer,
+            fsm_grammar_states=fsm_grammar_states,
         )
 
         speculative_ids = (

server/text_generation_server/utils/tokens.py

@@ -466,6 +466,7 @@ class HeterogeneousNextTokenChooser:
         dtype: torch.dtype,
         device: torch.device,
         tokenizer: PreTrainedTokenizerBase,
+        fsm_grammar_states: Optional[List[int]] = None,
     ) -> "HeterogeneousNextTokenChooser":
         return HeterogeneousNextTokenChooser(
             watermark=[pb_.watermark for pb_ in pb],
@@ -482,7 +483,9 @@ class HeterogeneousNextTokenChooser:
             tokenizer=tokenizer,
             grammars=[pb_.grammar for pb_ in pb],
             grammar_types=[pb_.grammar_type for pb_ in pb],
-            fsm_grammar_states=[0] * len(pb),
+            fsm_grammar_states=(
+                fsm_grammar_states if fsm_grammar_states else [0] * len(pb)
+            ),
         )