[Model] Adding support for MiniCPM-V (#4087)

This commit is contained in:
Alphi
2024-07-24 20:59:30 -07:00
committed by GitHub
parent 5689e256ba
commit 9e169a4c61
11 changed files with 942 additions and 18 deletions
+6 -5
View File
@@ -11,7 +11,7 @@ import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from transformers import (AutoModelForCausalLM, AutoModelForVision2Seq,
AutoTokenizer, BatchEncoding)
AutoTokenizer, BatchEncoding, BatchFeature)
from vllm import LLM, SamplingParams
from vllm.assets.image import ImageAsset
@@ -133,7 +133,7 @@ def image_assets() -> _ImageAssets:
return IMAGE_ASSETS
_T = TypeVar("_T", nn.Module, torch.Tensor, BatchEncoding)
_T = TypeVar("_T", nn.Module, torch.Tensor, BatchEncoding, BatchFeature)
class HfRunner:
@@ -339,7 +339,6 @@ class HfRunner:
processor_kwargs["images"] = images[i]
inputs = self.processor(**processor_kwargs)
input_ids = inputs.input_ids
output = self.model.generate(
**self.wrap_device(inputs),
@@ -381,7 +380,7 @@ class HfRunner:
all_logprobs.append(seq_logprobs_lst)
seq_ids = output.sequences[0]
output_len = seq_ids.shape[0] - input_ids.shape[1]
output_len = len(seq_logprobs_lst)
output_ids = seq_ids[-output_len:]
all_output_ids.append(output_ids.tolist())
all_output_strs.append(self.tokenizer.decode(output_ids))
@@ -514,10 +513,12 @@ class VllmRunner:
max_tokens: int,
num_logprobs: int,
images: Optional[List[Image.Image]] = None,
stop_token_ids: Optional[List[int]] = None,
) -> List[Tuple[List[int], str, Optional[SampleLogprobs]]]:
greedy_logprobs_params = SamplingParams(temperature=0.0,
max_tokens=max_tokens,
logprobs=num_logprobs)
logprobs=num_logprobs,
stop_token_ids=stop_token_ids)
outputs = self.generate_w_logprobs(prompts,
greedy_logprobs_params,
images=images)