Mirrors
/
hf_text-generation-inference
mirror of https://github.com/huggingface/text-generation-inference.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

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
This commit is contained in:
drbh 2024-02-29 05:17:42 -05:00 committed by GitHub
parent e6bb3ff81f
commit 343aa7a197
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 11 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
integration-tests/models/__snapshots__/test_grammar_llama/test_flash_llama_grammar_load.json
View File

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

3
server/text_generation_server/models/flash_causal_lm.py
View File

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

5
server/text_generation_server/utils/tokens.py
View File

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