Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Generate: handle logits_warper update in models with custom generate fn #31957

Merged
merged 1 commit into from
Jul 15, 2024
Merged

Generate: handle logits_warper update in models with custom generate fn #31957

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions src/transformers/generation/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2219,7 +2219,7 @@ def _dola_decoding(
generation_config: GenerationConfig,
synced_gpus: bool,
streamer: "BaseStreamer",
logits_warper: LogitsProcessorList,
logits_warper: Optional[LogitsProcessorList],
**model_kwargs,
) -> Union[GenerateNonBeamOutput, torch.LongTensor]:
r"""
Expand Down Expand Up @@ -2826,7 +2826,7 @@ def _sample(
generation_config: GenerationConfig,
synced_gpus: bool,
streamer: Optional["BaseStreamer"],
logits_warper: LogitsProcessorList,
logits_warper: Optional[LogitsProcessorList],
**model_kwargs,
) -> Union[GenerateNonBeamOutput, torch.LongTensor]:
r"""
Expand Down Expand Up @@ -3033,7 +3033,7 @@ def _beam_search(
stopping_criteria: StoppingCriteriaList,
generation_config: GenerationConfig,
synced_gpus: bool,
logits_warper: LogitsProcessorList,
logits_warper: Optional[LogitsProcessorList],
**model_kwargs,
) -> Union[GenerateBeamOutput, torch.LongTensor]:
r"""
Expand Down
80 changes: 17 additions & 63 deletions src/transformers/models/musicgen/modeling_musicgen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...generation.configuration_utils import GenerationConfig
from ...generation.configuration_utils import GenerationConfig, GenerationMode
from ...generation.logits_process import ClassifierFreeGuidanceLogitsProcessor, LogitsProcessorList
from ...generation.stopping_criteria import StoppingCriteriaList
from ...modeling_attn_mask_utils import (
Expand Down Expand Up @@ -1618,16 +1618,7 @@ def generate(
model_kwargs["delay_pattern_mask"] = delay_pattern_mask

# 7. determine generation mode
is_greedy_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is False
)
is_sample_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is True
)
generation_mode = generation_config.get_generation_mode()

# 8. prepare batched CFG externally (to enable coexistance with the unbatched CFG)
if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
Expand All @@ -1649,27 +1640,13 @@ def generate(
generation_config=generation_config, stopping_criteria=stopping_criteria
)

if is_greedy_gen_mode:
if generation_config.num_return_sequences > 1:
raise ValueError(
"num_return_sequences has to be 1 when doing greedy search, "
f"but is {generation_config.num_return_sequences}."
)

# 11. run greedy search
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
streamer=streamer,
**model_kwargs,
)

elif is_sample_gen_mode:
if generation_mode in (GenerationMode.SAMPLE, GenerationMode.GREEDY_SEARCH):
# 11. prepare logits warper
logits_warper = self._get_logits_warper(generation_config, device=input_ids.device)
prepared_logits_warper = (
self._get_logits_warper(generation_config, device=input_ids.device)
if generation_config.do_sample
else None
)

# expand input_ids with `num_return_sequences` additional sequences per batch
input_ids, model_kwargs = self._expand_inputs_for_generation(
Expand All @@ -1682,7 +1659,7 @@ def generate(
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper=prepared_logits_warper,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
Expand Down Expand Up @@ -2714,16 +2691,7 @@ def generate(
streamer.put(input_ids.cpu())

# 7. determine generation mode
is_greedy_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is False
)
is_sample_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is True
)
generation_mode = generation_config.get_generation_mode()

# 8. prepare batched CFG externally (to enable coexistance with the unbatched CFG)
if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
Expand All @@ -2745,27 +2713,13 @@ def generate(
generation_config=generation_config, stopping_criteria=stopping_criteria
)

if is_greedy_gen_mode:
if generation_config.num_return_sequences > 1:
raise ValueError(
"num_return_sequences has to be 1 when doing greedy search, "
f"but is {generation_config.num_return_sequences}."
)

# 11. run greedy search
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
streamer=streamer,
**model_kwargs,
)

elif is_sample_gen_mode:
if generation_mode in (GenerationMode.SAMPLE, GenerationMode.GREEDY_SEARCH):
# 11. prepare logits warper
logits_warper = self._get_logits_warper(generation_config, device=input_ids.device)
prepared_logits_warper = (
self._get_logits_warper(generation_config, device=input_ids.device)
if generation_config.do_sample
else None
)

# expand input_ids with `num_return_sequences` additional sequences per batch
input_ids, model_kwargs = self._expand_inputs_for_generation(
Expand All @@ -2779,7 +2733,7 @@ def generate(
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper=prepared_logits_warper,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
Expand Down
80 changes: 17 additions & 63 deletions src/transformers/models/musicgen_melody/modeling_musicgen_melody.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@
from torch.nn import CrossEntropyLoss

from ...activations import ACT2FN
from ...generation.configuration_utils import GenerationConfig
from ...generation.configuration_utils import GenerationConfig, GenerationMode
from ...generation.logits_process import ClassifierFreeGuidanceLogitsProcessor, LogitsProcessorList
from ...generation.stopping_criteria import StoppingCriteriaList
from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
Expand Down Expand Up @@ -1539,16 +1539,7 @@ def generate(
model_kwargs["delay_pattern_mask"] = delay_pattern_mask

# 7. determine generation mode
is_greedy_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is False
)
is_sample_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is True
)
generation_mode = generation_config.get_generation_mode()

# 8. prepare batched CFG externally (to enable coexistance with the unbatched CFG)
if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
Expand All @@ -1570,27 +1561,13 @@ def generate(
generation_config=generation_config, stopping_criteria=stopping_criteria
)

if is_greedy_gen_mode:
if generation_config.num_return_sequences > 1:
raise ValueError(
"num_return_sequences has to be 1 when doing greedy search, "
f"but is {generation_config.num_return_sequences}."
)

# 11. run greedy search
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
streamer=streamer,
**model_kwargs,
)

elif is_sample_gen_mode:
if generation_mode in (GenerationMode.SAMPLE, GenerationMode.GREEDY_SEARCH):
# 11. prepare logits warper
logits_warper = self._get_logits_warper(generation_config, device=input_ids.device)
prepared_logits_warper = (
self._get_logits_warper(generation_config, device=input_ids.device)
if generation_config.do_sample
else None
)

# expand input_ids with `num_return_sequences` additional sequences per batch
input_ids, model_kwargs = self._expand_inputs_for_generation(
Expand All @@ -1603,7 +1580,7 @@ def generate(
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper=prepared_logits_warper,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
Expand Down Expand Up @@ -2557,16 +2534,7 @@ def generate(
streamer.put(input_ids.cpu())

# 7. determine generation mode
is_greedy_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is False
)
is_sample_gen_mode = (
(generation_config.num_beams == 1)
and (generation_config.num_beam_groups == 1)
and generation_config.do_sample is True
)
generation_mode = generation_config.get_generation_mode()

# 8. prepare batched CFG externally (to enable coexistance with the unbatched CFG)
if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
Expand All @@ -2588,27 +2556,13 @@ def generate(
generation_config=generation_config, stopping_criteria=stopping_criteria
)

if is_greedy_gen_mode:
if generation_config.num_return_sequences > 1:
raise ValueError(
"num_return_sequences has to be 1 when doing greedy search, "
f"but is {generation_config.num_return_sequences}."
)

# 11. run greedy search
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
streamer=streamer,
**model_kwargs,
)

elif is_sample_gen_mode:
if generation_mode in (GenerationMode.SAMPLE, GenerationMode.GREEDY_SEARCH):
# 11. prepare logits warper
logits_warper = self._get_logits_warper(generation_config, device=input_ids.device)
prepared_logits_warper = (
self._get_logits_warper(generation_config, device=input_ids.device)
if generation_config.do_sample
else None
)

# expand input_ids with `num_return_sequences` additional sequences per batch
input_ids, model_kwargs = self._expand_inputs_for_generation(
Expand All @@ -2622,7 +2576,7 @@ def generate(
outputs = self._sample(
input_ids,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper=prepared_logits_warper,
stopping_criteria=stopping_criteria,
generation_config=generation_config,
synced_gpus=synced_gpus,
Expand Down
2 changes: 2 additions & 0 deletions src/transformers/models/rag/modeling_rag.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1558,6 +1558,7 @@ def extend_enc_output(tensor, num_beams=None):
generation_config=generation_config,
synced_gpus=False,
streamer=None,
logits_warper=None,
**model_kwargs,
)
elif generation_config.num_beams > 1:
Expand All @@ -1579,6 +1580,7 @@ def extend_enc_output(tensor, num_beams=None):
stopping_criteria=prepared_stopping_criteria,
generation_config=generation_config,
synced_gpus=False,
logits_warper=None,
**model_kwargs,
)
else:
Expand Down
Loading