Support Cross encoder models (#10400)

Signed-off-by: Max de Bayser <maxdebayser@gmail.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Signed-off-by: Flavia Beo <flavia.beo@ibm.com>
Co-authored-by: Flavia Beo <flavia.beo@ibm.com>
This commit is contained in:
Maximilien de Bayser
2024-11-24 18:56:20 -08:00
committed by GitHub
co-authored by Flavia Beo
parent 49628fe13e
commit 214efc2c3c
28 changed files with 1370 additions and 62 deletions
+28
View File
@@ -92,6 +92,7 @@ class ModelInputForGPU(ModelRunnerInputBase):
"""
input_tokens: Optional[torch.Tensor] = None
input_positions: Optional[torch.Tensor] = None
token_types: Optional[torch.Tensor] = None
seq_lens: Optional[List[int]] = None
query_lens: Optional[List[int]] = None
lora_mapping: Optional["LoRAMapping"] = None
@@ -200,6 +201,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
def simple_reinit(self):
self.input_tokens[0].clear() # type: ignore
self.input_positions[0].clear() # type: ignore
self.token_types[0].clear() # type: ignore
self.mrope_input_positions = None # type: ignore
self.seq_lens[0] = 0 # type: ignore
self.orig_seq_lens[0] = 0 # type: ignore
@@ -226,6 +228,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
# Input tokens and positions.
input_tokens: Optional[List[List[int]]] = None,
input_positions: Optional[List[List[int]]] = None,
token_types: Optional[List[List[int]]] = None,
mrope_input_positions: Optional[List[List[List[int]]]] = None,
# The sequence length (may be capped to the sliding window).
@@ -291,6 +294,12 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
for seq_id in range(len(self.seq_ids)):
self.input_positions[seq_id].clear()
if token_types:
self.token_types = token_types
else:
for seq_id in range(len(self.seq_ids)):
self.token_types[seq_id].clear()
self.mrope_input_positions = None
if seq_lens:
@@ -354,6 +363,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
else:
self.input_tokens = input_tokens or []
self.input_positions = input_positions or []
self.token_types = token_types or []
self.mrope_input_positions = mrope_input_positions or None
self.seq_lens = seq_lens or []
self.orig_seq_lens = orig_seq_lens or []
@@ -386,6 +396,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
self.input_tokens = [[] for _ in range(self.n_seqs)]
self.input_positions = [[] for _ in range(self.n_seqs)]
self.token_types = [[] for _ in range(self.n_seqs)]
self.mrope_input_positions = None
self.seq_lens = [0] * self.n_seqs
self.orig_seq_lens = [0] * self.n_seqs
@@ -498,12 +509,15 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
# Compute tokens.
tokens = seq_data.get_token_ids()[context_len:seq_len]
token_types = seq_group_metadata.token_type_ids
inter_data.seq_lens[seq_idx] = seq_len
inter_data.orig_seq_lens[seq_idx] = seq_len
inter_data.context_lens[seq_idx] = context_len
inter_data.input_tokens[seq_idx].extend(tokens)
inter_data.input_positions[seq_idx].extend(range(context_len, seq_len))
inter_data.token_types[seq_idx].extend(
token_types if token_types else [])
inter_data.query_lens[seq_idx] = seq_len - context_len
if seq_data.mrope_position_delta is not None:
@@ -561,6 +575,8 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
seq_idx][uncomputed_start:]
inter_data.input_positions[seq_idx] = inter_data.input_positions[
seq_idx][uncomputed_start:]
inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
uncomputed_start:]
context_len = prefix_cache_len
inter_data.context_lens[seq_idx] = context_len
@@ -575,6 +591,8 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
seq_idx][-1:]
inter_data.input_positions[seq_idx] = inter_data.input_positions[
seq_idx][-1:]
inter_data.token_types[seq_idx] = inter_data.token_types[seq_idx][
-1:]
inter_data.query_lens[seq_idx] = 1
inter_data.context_lens[seq_idx] = inter_data.seq_lens[seq_idx] - 1
@@ -803,9 +821,12 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
"""
# Combine and flatten intermediate data.
input_tokens = []
token_types = []
for inter_data in self.inter_data_list:
for cur_input_tokens in inter_data.input_tokens:
input_tokens.extend(cur_input_tokens)
for cur_token_types in inter_data.token_types:
token_types.extend(cur_token_types)
if not input_tokens:
# This may happen when all prefill requests hit
@@ -874,6 +895,12 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
input_tokens_tensor = async_tensor_h2d(input_tokens, torch.long,
self.runner.device,
self.runner.pin_memory)
token_types_tensor = async_tensor_h2d(token_types, torch.long,
self.runner.device,
self.runner.pin_memory) \
if token_types else None
if mrope_input_positions is not None:
for idx in range(3):
mrope_input_positions[idx].extend(
@@ -952,6 +979,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
return self.model_input_cls(
input_tokens=input_tokens_tensor,
input_positions=input_positions_tensor,
token_types=token_types_tensor,
attn_metadata=attn_metadata,
seq_lens=seq_lens,
query_lens=query_lens,