[Core] Automatically cast multi-modal input dtype (#18756)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-05-27 23:45:48 +08:00
committed by GitHub
parent 6b6d496114
commit 696259ca01
16 changed files with 91 additions and 44 deletions
+5 -2
View File
@@ -297,8 +297,11 @@ class CPUEncoderDecoderModelRunner(
model_input.encoder_input_tokens,
"encoder_positions":
model_input.encoder_input_positions,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
"intermediate_tensors":
intermediate_tensors,
}
+4 -1
View File
@@ -628,7 +628,10 @@ class CPUModelRunner(CPUModelRunnerBase[ModelInputForCPUWithSamplingMetadata]):
multimodal_kwargs = {}
if model_input.multi_modal_kwargs is not None:
multimodal_kwargs = MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs, device=self.device)
model_input.multi_modal_kwargs,
dtype=self.model_config.dtype,
device=self.device,
)
execute_model_kwargs = {}
if previous_hidden_states is not None:
execute_model_kwargs.update(
+5 -2
View File
@@ -50,8 +50,11 @@ class CPUPoolingModelRunner(
model_input.input_tokens,
"positions":
model_input.input_positions,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
**cross_enc_kwargs,
"intermediate_tensors":
intermediate_tensors,
+7 -3
View File
@@ -202,9 +202,13 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
encoder_input_ids=model_input.encoder_input_tokens,
encoder_positions=model_input.encoder_input_positions,
intermediate_tensors=intermediate_tensors,
**MultiModalKwargs.as_kwargs(multi_modal_kwargs,
device=self.device),
**seqlen_agnostic_kwargs)
**MultiModalKwargs.as_kwargs(
multi_modal_kwargs,
dtype=self.model_config.dtype,
device=self.device,
),
**seqlen_agnostic_kwargs,
)
logits = self.model.compute_logits(hidden_or_intermediate_states,
model_input.sampling_metadata)
+5 -2
View File
@@ -1845,8 +1845,11 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
inputs_embeds=model_input.inputs_embeds,
positions=model_input.input_positions,
intermediate_tensors=intermediate_tensors,
**MultiModalKwargs.as_kwargs(multi_modal_kwargs,
device=self.device),
**MultiModalKwargs.as_kwargs(
multi_modal_kwargs,
dtype=self.model_config.dtype,
device=self.device,
),
**seqlen_agnostic_kwargs,
**model_kwargs,
)
@@ -70,8 +70,11 @@ class MultiStepNeuronModelRunner(NeuronModelRunner):
input_ids=model_input.input_tokens,
positions=model_input.input_positions,
input_block_ids=model_input.input_block_ids,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
)
output = self.model.sample(
@@ -49,8 +49,11 @@ class MultiStepNeuronxDistributedModelRunner(NeuronxDistributedModelRunner):
positions=model_input.input_positions,
input_block_ids=model_input.input_block_ids,
sampling_params=sampling_params,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
)
output = self.model.sample(
+10 -6
View File
@@ -378,9 +378,11 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
positions=model_input.input_positions,
input_block_ids=model_input.input_block_ids,
sampling_params=sampling_params,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
)
elif current_platform.use_transformers_neuronx():
# [TODO] validate on-device sampling
@@ -389,9 +391,11 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
input_ids=model_input.input_tokens,
positions=model_input.input_positions,
input_block_ids=model_input.input_block_ids,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
or {},
device=self.device),
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
)
# Compute the logits only if the on-device sampling is turned off as
+7 -3
View File
@@ -119,10 +119,14 @@ class PoolingModelRunner(
input_ids=model_input.input_tokens,
positions=model_input.input_positions,
intermediate_tensors=intermediate_tensors,
**MultiModalKwargs.as_kwargs(multi_modal_kwargs,
device=self.device),
**MultiModalKwargs.as_kwargs(
multi_modal_kwargs,
dtype=self.model_config.dtype,
device=self.device,
),
**cross_enc_kwargs,
**seqlen_agnostic_kwargs)
**seqlen_agnostic_kwargs,
)
if (self.observability_config is not None
and self.observability_config.collect_model_forward_time):
+6 -3
View File
@@ -562,9 +562,12 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
input_ids=model_input.input_tokens,
positions=model_input.input_positions,
intermediate_tensors=intermediate_tensors,
**MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs
or {},
device=self.device))
**MultiModalKwargs.as_kwargs(
model_input.multi_modal_kwargs or {},
dtype=self.model_config.dtype,
device=self.device,
),
)
# Compute the logits in the last pipeline stage.
if not get_pp_group().is_last_rank:
return hidden_or_intermediate_states