Add files using upload-large-folder tool
Browse files- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/__init__.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/config.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/detokenizer.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/detokenizer_utils.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/dynamic_module.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/processor.cpython-312.pyc +0 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/mistral.py +164 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/mlp_speculator.py +68 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/moonvit.py +33 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron.py +205 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron_h.py +259 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron_vl.py +56 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/ovis.py +176 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/step3_vl.py +123 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/ultravox.py +112 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/__init__.py +15 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/deepseek_vl2.py +363 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/ovis.py +420 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/tokenizers/__init__.py +10 -0
- tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/tokenizers/mistral.py +516 -0
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/__init__.cpython-312.pyc
ADDED
|
Binary file (962 Bytes). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/config.cpython-312.pyc
ADDED
|
Binary file (32.7 kB). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/detokenizer.cpython-312.pyc
ADDED
|
Binary file (5.79 kB). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/detokenizer_utils.cpython-312.pyc
ADDED
|
Binary file (5.89 kB). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/dynamic_module.cpython-312.pyc
ADDED
|
Binary file (1.93 kB). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/__pycache__/processor.cpython-312.pyc
ADDED
|
Binary file (7.93 kB). View file
|
|
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/mistral.py
ADDED
|
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
from typing import Any
|
| 4 |
+
|
| 5 |
+
from transformers import PretrainedConfig, WhisperConfig
|
| 6 |
+
|
| 7 |
+
from vllm.logger import init_logger
|
| 8 |
+
|
| 9 |
+
logger = init_logger(__name__)
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
def adapt_config_dict(config_dict: dict[str, Any],
|
| 13 |
+
**kwargs) -> PretrainedConfig:
|
| 14 |
+
config_dict.update(kwargs)
|
| 15 |
+
config_dict = _remap_general_mistral_args(config_dict)
|
| 16 |
+
|
| 17 |
+
if bool(config_dict.get("quantization")):
|
| 18 |
+
config_dict = _remap_mistral_quantization_args(config_dict)
|
| 19 |
+
|
| 20 |
+
if bool(config_dict.get("moe")):
|
| 21 |
+
config_dict["architectures"] = ["MixtralForCausalLM"]
|
| 22 |
+
else:
|
| 23 |
+
config_dict["architectures"] = ["MistralForCausalLM"]
|
| 24 |
+
|
| 25 |
+
if bool(config_dict.get("yarn")):
|
| 26 |
+
config_dict = _remap_mistral_yarn_args(config_dict)
|
| 27 |
+
|
| 28 |
+
is_vision = ((config_dict.get("multimodal")
|
| 29 |
+
or {}).get("vision_encoder_args")
|
| 30 |
+
or config_dict.get("vision_encoder"))
|
| 31 |
+
is_audio = bool(
|
| 32 |
+
((config_dict.get("multimodal") or {}).get("whisper_model_args")
|
| 33 |
+
or {}).get("encoder_args"))
|
| 34 |
+
|
| 35 |
+
assert not (is_vision and is_audio), \
|
| 36 |
+
"Vision and audio are mutually exclusive"
|
| 37 |
+
|
| 38 |
+
if is_vision:
|
| 39 |
+
config_dict = _remap_mistral_vision_args(config_dict)
|
| 40 |
+
if is_audio:
|
| 41 |
+
config_dict = _remap_mistral_audio_args(config_dict)
|
| 42 |
+
|
| 43 |
+
config = PretrainedConfig.from_dict(config_dict)
|
| 44 |
+
|
| 45 |
+
logger.debug("Initialized config %s", config)
|
| 46 |
+
|
| 47 |
+
return config
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def _remap_mistral_vision_args(config: dict) -> dict:
|
| 51 |
+
if config.get("multimodal"):
|
| 52 |
+
vision_config = config.pop("multimodal")
|
| 53 |
+
else:
|
| 54 |
+
vision_config = config.pop("vision_encoder")
|
| 55 |
+
|
| 56 |
+
quant_config = config.get("quantization_config")
|
| 57 |
+
config = {
|
| 58 |
+
"model_type": "pixtral",
|
| 59 |
+
"architectures": ["PixtralForConditionalGeneration"],
|
| 60 |
+
"text_config": PretrainedConfig.from_dict(config),
|
| 61 |
+
"vision_config": PretrainedConfig.from_dict(vision_config),
|
| 62 |
+
}
|
| 63 |
+
if quant_config:
|
| 64 |
+
config["quantization_config"] = quant_config
|
| 65 |
+
return config
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def _remap_mistral_yarn_args(config: dict) -> dict:
|
| 69 |
+
# Direct remaps: yarn.X -> rope_scaling.Y
|
| 70 |
+
# Source keys are from mistral.model.args.YarnArgs
|
| 71 |
+
_map = {
|
| 72 |
+
"beta": "beta_fast",
|
| 73 |
+
"alpha": "beta_slow",
|
| 74 |
+
}
|
| 75 |
+
yarn_config = config.get("yarn") or {}
|
| 76 |
+
renamed_yarn_config = {_map.get(k, k): v for k, v in yarn_config.items()}
|
| 77 |
+
config["rope_scaling"] = {
|
| 78 |
+
"rope_type": "yarn",
|
| 79 |
+
"mscale_all_dim": 1, # We hardcoded this to 1
|
| 80 |
+
**renamed_yarn_config
|
| 81 |
+
}
|
| 82 |
+
return config
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
def _remap_general_mistral_args(config: dict) -> dict:
|
| 86 |
+
# Mistral key -> HF key
|
| 87 |
+
config_mapping = {
|
| 88 |
+
"dim": "hidden_size",
|
| 89 |
+
"norm_eps": "rms_norm_eps",
|
| 90 |
+
"n_kv_heads": "num_key_value_heads",
|
| 91 |
+
"n_layers": "num_hidden_layers",
|
| 92 |
+
"n_heads": "num_attention_heads",
|
| 93 |
+
"hidden_dim": "intermediate_size",
|
| 94 |
+
}
|
| 95 |
+
# HF key -> (Mistral key, default value)
|
| 96 |
+
top_level_mapping_with_default = {
|
| 97 |
+
"model_type": ("model_type", "transformer"),
|
| 98 |
+
"hidden_act": ("activation", "silu"),
|
| 99 |
+
"tie_word_embeddings": ("tied_embeddings", False),
|
| 100 |
+
"max_seq_len": ("max_seq_len", 128_000),
|
| 101 |
+
"max_position_embeddings": ("max_position_embeddings", 128_000),
|
| 102 |
+
}
|
| 103 |
+
|
| 104 |
+
for key, new_key in config_mapping.items():
|
| 105 |
+
if key in config:
|
| 106 |
+
config[new_key] = config.pop(key)
|
| 107 |
+
|
| 108 |
+
for new_key, (key,
|
| 109 |
+
default_value) in top_level_mapping_with_default.items():
|
| 110 |
+
config[new_key] = config.pop(key, default_value)
|
| 111 |
+
|
| 112 |
+
return config
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def _remap_mistral_quantization_args(config: dict) -> dict:
|
| 116 |
+
quantization = config.get("quantization", {})
|
| 117 |
+
if quantization.get("qformat_weight") == "fp8_e4m3":
|
| 118 |
+
# This maps to the FP8 static per-tensor quantization scheme
|
| 119 |
+
quantization_config = {
|
| 120 |
+
"quant_method": "fp8",
|
| 121 |
+
"activation_scheme": "static"
|
| 122 |
+
}
|
| 123 |
+
elif quantization.get("quant_method") == "compressed-tensors":
|
| 124 |
+
# Pass through the quantization config to compressed-tensors
|
| 125 |
+
quantization_config = quantization
|
| 126 |
+
else:
|
| 127 |
+
raise ValueError(
|
| 128 |
+
f"Found unknown quantization='{quantization}' in config")
|
| 129 |
+
|
| 130 |
+
config["quantization_config"] = quantization_config
|
| 131 |
+
|
| 132 |
+
return config
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
def _remap_mistral_audio_args(config: dict) -> dict:
|
| 136 |
+
whisper_args = config["multimodal"].pop("whisper_model_args")
|
| 137 |
+
encoder_args = whisper_args["encoder_args"]
|
| 138 |
+
downsample_args = whisper_args["downsample_args"]
|
| 139 |
+
|
| 140 |
+
quant_config = config.get("quantization_config")
|
| 141 |
+
config = {
|
| 142 |
+
"model_type":
|
| 143 |
+
"whixtral",
|
| 144 |
+
"architectures": ["VoxtralForConditionalGeneration"],
|
| 145 |
+
"text_config":
|
| 146 |
+
PretrainedConfig.from_dict(config),
|
| 147 |
+
"audio_config":
|
| 148 |
+
WhisperConfig(
|
| 149 |
+
num_mel_bins=encoder_args["audio_encoding_args"]["num_mel_bins"],
|
| 150 |
+
window_size=encoder_args["audio_encoding_args"]["window_size"],
|
| 151 |
+
sampling_rate=encoder_args["audio_encoding_args"]["sampling_rate"],
|
| 152 |
+
hop_length=encoder_args["audio_encoding_args"]["hop_length"],
|
| 153 |
+
downsample_factor=downsample_args["downsample_factor"],
|
| 154 |
+
d_model=encoder_args["dim"],
|
| 155 |
+
encoder_layers=encoder_args["n_layers"],
|
| 156 |
+
encoder_ffn_dim=encoder_args["hidden_dim"],
|
| 157 |
+
encoder_attention_heads=encoder_args["n_heads"],
|
| 158 |
+
vocab_size=encoder_args["vocab_size"],
|
| 159 |
+
max_source_positions=encoder_args["max_source_positions"],
|
| 160 |
+
)
|
| 161 |
+
}
|
| 162 |
+
if quant_config:
|
| 163 |
+
config["quantization_config"] = quant_config
|
| 164 |
+
return config
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/mlp_speculator.py
ADDED
|
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
from typing import Optional
|
| 5 |
+
|
| 6 |
+
from transformers import PretrainedConfig
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class MLPSpeculatorConfig(PretrainedConfig):
|
| 10 |
+
model_type = "mlp_speculator"
|
| 11 |
+
|
| 12 |
+
attribute_map = {
|
| 13 |
+
"hidden_size": "emb_dim",
|
| 14 |
+
}
|
| 15 |
+
|
| 16 |
+
def __init__(self,
|
| 17 |
+
vocab_size: int = 32000,
|
| 18 |
+
emb_dim: int = 4096,
|
| 19 |
+
inner_dim: int = 0,
|
| 20 |
+
n_predict: int = 3,
|
| 21 |
+
top_k_tokens_per_head: Optional[list[int]] = None,
|
| 22 |
+
n_candidates: int = 5,
|
| 23 |
+
tie_weights: bool = False,
|
| 24 |
+
scale_input: bool = False,
|
| 25 |
+
**kwargs):
|
| 26 |
+
"""
|
| 27 |
+
Initialize an MLPSpeculatorConfig
|
| 28 |
+
|
| 29 |
+
Args:
|
| 30 |
+
vocab_size: int
|
| 31 |
+
the model vocab size
|
| 32 |
+
emb_dim: int
|
| 33 |
+
the model embedding dimension
|
| 34 |
+
inner_dim: int
|
| 35 |
+
the inner dimension of the model. If 0, will be the emb_dim.
|
| 36 |
+
n_predict: int
|
| 37 |
+
the number of lookaheads for the speculator
|
| 38 |
+
top_k_tokens_per_head: list[int]
|
| 39 |
+
Number of tokens to consider from each head when forming the
|
| 40 |
+
candidate tree.
|
| 41 |
+
For each candidate branch in the tree, head n produces topk[n]
|
| 42 |
+
additional sub-branches.
|
| 43 |
+
NOTE: This parameter is currently unused.
|
| 44 |
+
n_candidates: int
|
| 45 |
+
number of child candidates to create per sequence
|
| 46 |
+
tie_weights: bool
|
| 47 |
+
If true, use a single set of weights for every model
|
| 48 |
+
head/stage after the first. The initial projection
|
| 49 |
+
from the base model may have a different size, so that
|
| 50 |
+
stays separate.
|
| 51 |
+
scale_input: bool
|
| 52 |
+
if True, will scale the initial hidden states from
|
| 53 |
+
the base model.
|
| 54 |
+
"""
|
| 55 |
+
if top_k_tokens_per_head is None:
|
| 56 |
+
top_k_tokens_per_head = [5, 4, 3]
|
| 57 |
+
assert len(top_k_tokens_per_head) == n_predict
|
| 58 |
+
self.vocab_size = vocab_size
|
| 59 |
+
self.emb_dim = emb_dim
|
| 60 |
+
self.inner_dim = inner_dim
|
| 61 |
+
self.n_predict = n_predict
|
| 62 |
+
self.top_k_tokens_per_head = top_k_tokens_per_head
|
| 63 |
+
self.n_candidates = n_candidates
|
| 64 |
+
self.num_lookahead_tokens = n_predict
|
| 65 |
+
self.tie_weights = tie_weights
|
| 66 |
+
self.scale_input = scale_input
|
| 67 |
+
|
| 68 |
+
super().__init__(**kwargs)
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/moonvit.py
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
|
| 4 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
class MoonViTConfig(PretrainedConfig):
|
| 8 |
+
model_type = "moonvit"
|
| 9 |
+
|
| 10 |
+
def __init__(
|
| 11 |
+
self,
|
| 12 |
+
patch_size: int = 14,
|
| 13 |
+
init_pos_emb_height: int = 64,
|
| 14 |
+
init_pos_emb_width: int = 64,
|
| 15 |
+
num_attention_heads: int = 16,
|
| 16 |
+
num_hidden_layers: int = 27,
|
| 17 |
+
hidden_size: int = 1152,
|
| 18 |
+
intermediate_size: int = 4304,
|
| 19 |
+
merge_kernel_size: tuple[int, int] = (2, 2),
|
| 20 |
+
**kwargs,
|
| 21 |
+
):
|
| 22 |
+
super().__init__(**kwargs)
|
| 23 |
+
self.patch_size = patch_size
|
| 24 |
+
# Positional embedding config
|
| 25 |
+
self.init_pos_emb_height = init_pos_emb_height
|
| 26 |
+
self.init_pos_emb_width = init_pos_emb_width
|
| 27 |
+
# Transformer config
|
| 28 |
+
self.num_hidden_layers = num_hidden_layers
|
| 29 |
+
self.num_attention_heads = num_attention_heads
|
| 30 |
+
self.hidden_size = hidden_size
|
| 31 |
+
self.intermediate_size = intermediate_size
|
| 32 |
+
# Patch merger config
|
| 33 |
+
self.merge_kernel_size = merge_kernel_size
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron.py
ADDED
|
@@ -0,0 +1,205 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# Copyright 2024 HuggingFace Inc. team. All rights reserved.
|
| 5 |
+
# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
|
| 6 |
+
#
|
| 7 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 8 |
+
# you may not use this file except in compliance with the License.
|
| 9 |
+
# You may obtain a copy of the License at
|
| 10 |
+
#
|
| 11 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 12 |
+
#
|
| 13 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 14 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 15 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 16 |
+
# See the License for the specific language governing permissions and
|
| 17 |
+
# limitations under the License.
|
| 18 |
+
"""Nemotron model configuration"""
|
| 19 |
+
|
| 20 |
+
from transformers import PretrainedConfig
|
| 21 |
+
from transformers.utils import logging
|
| 22 |
+
|
| 23 |
+
logger = logging.get_logger(__name__)
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
class NemotronConfig(PretrainedConfig):
|
| 27 |
+
r"""
|
| 28 |
+
This is the configuration class to store the configuration of a
|
| 29 |
+
[`NemotronModel`]. It is used to instantiate an Nemotron model
|
| 30 |
+
according to the specified arguments, defining the model architecture.
|
| 31 |
+
Instantiating a configuration with the defaults will yield a similar
|
| 32 |
+
configuration to that of the Nemotron-8B.
|
| 33 |
+
|
| 34 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be
|
| 35 |
+
used to control the model outputs. Read the documentation from
|
| 36 |
+
[`PretrainedConfig`] for more information.
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
Args:
|
| 40 |
+
vocab_size (`int`, *optional*, defaults to 256000):
|
| 41 |
+
Vocabulary size of the Nemotron model. Defines the number of
|
| 42 |
+
different tokens that can be represented by the
|
| 43 |
+
`inputs_ids` passed when calling [`NemotronModel`]
|
| 44 |
+
hidden_size (`int`, *optional*, defaults to 6144):
|
| 45 |
+
Dimension of the hidden representations.
|
| 46 |
+
intermediate_size (`int`, *optional*, defaults to 24576):
|
| 47 |
+
Dimension of the MLP representations.
|
| 48 |
+
num_hidden_layers (`int`, *optional*, defaults to 32):
|
| 49 |
+
Number of hidden layers in the Transformer decoder.
|
| 50 |
+
num_attention_heads (`int`, *optional*, defaults to 48):
|
| 51 |
+
Number of attention heads for each attention layer in the
|
| 52 |
+
Transformer decoder.
|
| 53 |
+
head_dim (`int`, *optional*):
|
| 54 |
+
Projection weights dimension in multi-head attention. Set to
|
| 55 |
+
hidden_size // num_attention_heads if None
|
| 56 |
+
num_key_value_heads (`int`, *optional*):
|
| 57 |
+
This is the number of key_value heads that should be used to
|
| 58 |
+
implement Grouped Query Attention. If
|
| 59 |
+
`num_key_value_heads=num_attention_heads`, the model will use
|
| 60 |
+
Multi Head Attention (MHA), if
|
| 61 |
+
`num_key_value_heads=1 the model will use Multi Query Attention
|
| 62 |
+
(MQA) otherwise GQA is used. When converting a multi-head
|
| 63 |
+
checkpoint to a GQA checkpoint, each group key and value
|
| 64 |
+
head should be constructed by meanpooling all the original
|
| 65 |
+
heads within that group. For more details checkout
|
| 66 |
+
[this paper](https://arxiv.org/pdf/2305.13245.pdf). If it
|
| 67 |
+
is not specified, will default to `num_attention_heads`.
|
| 68 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"relu2"`):
|
| 69 |
+
The non-linear activation function (function or string) in the
|
| 70 |
+
decoder.
|
| 71 |
+
max_position_embeddings (`int`, *optional*, defaults to 4096):
|
| 72 |
+
The maximum sequence length that this model might ever be used
|
| 73 |
+
with.
|
| 74 |
+
initializer_range (`float`, *optional*, defaults to 0.0134):
|
| 75 |
+
The standard deviation of the truncated_normal_initializer for
|
| 76 |
+
initializing all weight matrices.
|
| 77 |
+
norm_eps (`float`, *optional*, defaults to 1e-05):
|
| 78 |
+
The epsilon used by the normalization layers.
|
| 79 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 80 |
+
Whether or not the model should return the last key/values
|
| 81 |
+
attentions (not used by all models). Only relevant if
|
| 82 |
+
`config.is_decoder=True`.
|
| 83 |
+
pad_token_id (`int`, *optional*):
|
| 84 |
+
Padding token id.
|
| 85 |
+
bos_token_id (`int`, *optional*, defaults to 2):
|
| 86 |
+
Beginning of stream token id.
|
| 87 |
+
eos_token_id (`int`, *optional*, defaults to 3):
|
| 88 |
+
End of stream token id.
|
| 89 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 90 |
+
Whether to tie weight embeddings
|
| 91 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 92 |
+
The base period of the RoPE embeddings.
|
| 93 |
+
partial_rotary_factor (`float`, *optional*, defaults to 0.5):
|
| 94 |
+
Percentage of the query and keys which will have rotary embedding.
|
| 95 |
+
attention_bias (`bool`, *optional*, defaults to `False`):
|
| 96 |
+
Whether to use a bias in the query, key, value and output
|
| 97 |
+
projection layers during self-attention.
|
| 98 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 99 |
+
The dropout ratio for the attention probabilities.
|
| 100 |
+
mlp_bias (`bool`, *optional*, defaults to `False`):
|
| 101 |
+
Whether to use a bias in up_proj and down_proj layers in the MLP
|
| 102 |
+
layers.
|
| 103 |
+
|
| 104 |
+
```python
|
| 105 |
+
>>> from transformers import NemotronModel, NemotronConfig
|
| 106 |
+
>>> # Initializing a Nemotron nemotron-15b style configuration
|
| 107 |
+
>>> configuration = NemotronConfig()
|
| 108 |
+
>>> # Initializing a model from the nemotron-15b style configuration
|
| 109 |
+
>>> model = NemotronModel(configuration)
|
| 110 |
+
>>> # Accessing the model configuration
|
| 111 |
+
>>> configuration = model.config
|
| 112 |
+
```"""
|
| 113 |
+
|
| 114 |
+
model_type = "nemotron"
|
| 115 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 116 |
+
|
| 117 |
+
def __init__(
|
| 118 |
+
self,
|
| 119 |
+
vocab_size=256000,
|
| 120 |
+
hidden_size=6144,
|
| 121 |
+
intermediate_size=24576,
|
| 122 |
+
num_hidden_layers=32,
|
| 123 |
+
num_attention_heads=48,
|
| 124 |
+
head_dim=None,
|
| 125 |
+
num_key_value_heads=None,
|
| 126 |
+
hidden_act="relu2",
|
| 127 |
+
max_position_embeddings=4096,
|
| 128 |
+
initializer_range=0.0134,
|
| 129 |
+
norm_eps=1e-5,
|
| 130 |
+
use_cache=True,
|
| 131 |
+
pad_token_id=None,
|
| 132 |
+
bos_token_id=2,
|
| 133 |
+
eos_token_id=3,
|
| 134 |
+
tie_word_embeddings=False,
|
| 135 |
+
rope_theta=10000.0,
|
| 136 |
+
rope_scaling=None,
|
| 137 |
+
partial_rotary_factor=0.5,
|
| 138 |
+
attention_bias=False,
|
| 139 |
+
attention_dropout=0.0,
|
| 140 |
+
mlp_bias=False,
|
| 141 |
+
**kwargs,
|
| 142 |
+
):
|
| 143 |
+
self.vocab_size = vocab_size
|
| 144 |
+
self.max_position_embeddings = max_position_embeddings
|
| 145 |
+
self.hidden_size = hidden_size
|
| 146 |
+
self.intermediate_size = intermediate_size
|
| 147 |
+
self.num_hidden_layers = num_hidden_layers
|
| 148 |
+
self.num_attention_heads = num_attention_heads
|
| 149 |
+
head_dim = head_dim or kwargs.get("kv_channels")
|
| 150 |
+
self.head_dim = head_dim if head_dim is not None else (
|
| 151 |
+
hidden_size // num_attention_heads)
|
| 152 |
+
|
| 153 |
+
# for backward compatibility
|
| 154 |
+
if num_key_value_heads is None:
|
| 155 |
+
num_key_value_heads = num_attention_heads
|
| 156 |
+
|
| 157 |
+
self.num_key_value_heads = num_key_value_heads
|
| 158 |
+
self.hidden_act = hidden_act
|
| 159 |
+
self.initializer_range = initializer_range
|
| 160 |
+
self.norm_eps = norm_eps
|
| 161 |
+
self.use_cache = use_cache
|
| 162 |
+
self.rope_theta = rope_theta
|
| 163 |
+
self.rope_scaling = rope_scaling
|
| 164 |
+
# for backward compatibility
|
| 165 |
+
partial_rotary_factor = kwargs.get("rope_percent") or kwargs.get(
|
| 166 |
+
"rope_percentage") or partial_rotary_factor
|
| 167 |
+
self.partial_rotary_factor = partial_rotary_factor
|
| 168 |
+
self._rope_scaling_validation()
|
| 169 |
+
self.attention_bias = attention_bias
|
| 170 |
+
self.attention_dropout = attention_dropout
|
| 171 |
+
self.mlp_bias = mlp_bias
|
| 172 |
+
|
| 173 |
+
super().__init__(
|
| 174 |
+
pad_token_id=pad_token_id,
|
| 175 |
+
bos_token_id=bos_token_id,
|
| 176 |
+
eos_token_id=eos_token_id,
|
| 177 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 178 |
+
**kwargs,
|
| 179 |
+
)
|
| 180 |
+
|
| 181 |
+
def _rope_scaling_validation(self):
|
| 182 |
+
"""
|
| 183 |
+
Validate the `rope_scaling` configuration.
|
| 184 |
+
"""
|
| 185 |
+
if self.rope_scaling is None:
|
| 186 |
+
return
|
| 187 |
+
|
| 188 |
+
if not isinstance(self.rope_scaling, dict) or len(
|
| 189 |
+
self.rope_scaling) != 2:
|
| 190 |
+
raise ValueError(
|
| 191 |
+
"`rope_scaling` must be a dictionary with two fields, "
|
| 192 |
+
f"`type` and `factor`, got {self.rope_scaling}")
|
| 193 |
+
rope_scaling_type = self.rope_scaling.get("type", None)
|
| 194 |
+
rope_scaling_factor = self.rope_scaling.get("factor", None)
|
| 195 |
+
if rope_scaling_type is None or rope_scaling_type not in [
|
| 196 |
+
"linear", "dynamic"
|
| 197 |
+
]:
|
| 198 |
+
raise ValueError(
|
| 199 |
+
"`rope_scaling`'s type field must be one of ['linear', "
|
| 200 |
+
f"'dynamic'], got {rope_scaling_type}")
|
| 201 |
+
if rope_scaling_factor is None or not isinstance(
|
| 202 |
+
rope_scaling_factor, float) or rope_scaling_factor <= 1.0:
|
| 203 |
+
raise ValueError(
|
| 204 |
+
"`rope_scaling`'s factor field must be a float > 1, got "
|
| 205 |
+
f"{rope_scaling_factor}")
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron_h.py
ADDED
|
@@ -0,0 +1,259 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# Copyright 2024 HuggingFace Inc. team. All rights reserved.
|
| 5 |
+
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
|
| 6 |
+
#
|
| 7 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 8 |
+
# you may not use this file except in compliance with the License.
|
| 9 |
+
# You may obtain a copy of the License at
|
| 10 |
+
#
|
| 11 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 12 |
+
#
|
| 13 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 14 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 15 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 16 |
+
# See the License for the specific language governing permissions and
|
| 17 |
+
# limitations under the License.
|
| 18 |
+
"""NemotronH model configuration"""
|
| 19 |
+
|
| 20 |
+
import regex as re
|
| 21 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 22 |
+
from transformers.utils import logging
|
| 23 |
+
|
| 24 |
+
logger = logging.get_logger(__name__)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class NemotronHConfig(PretrainedConfig):
|
| 28 |
+
r"""
|
| 29 |
+
This is the configuration class to store the configuration of a
|
| 30 |
+
[`NemotronHModel`]. It is used to instantiate a NemotronH model according
|
| 31 |
+
to the specified arguments, defining the model architecture. Instantiating
|
| 32 |
+
a configuration with the defaults will yield a similar configuration to
|
| 33 |
+
that of the NemotronH-v0.1 model.
|
| 34 |
+
Args:
|
| 35 |
+
vocab_size (`int`, *optional*, defaults to 131072):
|
| 36 |
+
Vocabulary size of the NemotronH model. Defines the number of
|
| 37 |
+
different tokens that can be represented by the `inputs_ids`
|
| 38 |
+
passed when calling [`NemotronHModel`]
|
| 39 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 40 |
+
Whether the model's input and output word embeddings should be
|
| 41 |
+
tied. Note that this is only relevant if the model has a output
|
| 42 |
+
word embedding layer.
|
| 43 |
+
hidden_size (`int`, *optional*, defaults to 4096):
|
| 44 |
+
Dimension of the hidden representations.
|
| 45 |
+
intermediate_size (`int`, *optional*, defaults to 21504):
|
| 46 |
+
Dimension of the MLP representations.
|
| 47 |
+
num_hidden_layers (`int`, *optional*, defaults to 52):
|
| 48 |
+
Number of hidden layers in the Transformer encoder.
|
| 49 |
+
hybrid_override_pattern (`str`, *optional*, defaults to
|
| 50 |
+
`"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
|
| 51 |
+
The pattern of the hybrid model. The pattern is a string of
|
| 52 |
+
characters where each character represents
|
| 53 |
+
M: Mamba2, *: Attention, -: MLP
|
| 54 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 55 |
+
Number of attention heads for each attention layer in the
|
| 56 |
+
Transformer encoder.
|
| 57 |
+
attention_head_dim (`int`, *optional*, defaults to 128):
|
| 58 |
+
Dimension of each attention head.
|
| 59 |
+
num_key_value_heads (`int`, *optional*, defaults to 8):
|
| 60 |
+
This is the number of key_value heads that should be used to
|
| 61 |
+
implement Grouped Query Attention. If
|
| 62 |
+
`num_key_value_heads=num_attention_heads`, the model will use
|
| 63 |
+
Multi Head Attention (MHA), if `num_key_value_heads=1` the model
|
| 64 |
+
will use Multi Query Attention (MQA) otherwise GQA is used.
|
| 65 |
+
mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
|
| 66 |
+
The non-linear activation function in the MLP layers.
|
| 67 |
+
attention_bias (`bool`, *optional*, defaults to `False`):
|
| 68 |
+
Whether to use bias in attention layers.
|
| 69 |
+
mlp_bias (`bool`, *optional*, defaults to `False`):
|
| 70 |
+
Whether to use bias in MLP layers.
|
| 71 |
+
use_bias (`bool`, *optional*, defaults to `False`):
|
| 72 |
+
Whether to use bias in the model.
|
| 73 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 74 |
+
The standard deviation of the truncated_normal_initializer for
|
| 75 |
+
initializing all weight matrices.
|
| 76 |
+
layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
|
| 77 |
+
The epsilon used by the layer normalization layers.
|
| 78 |
+
residual_in_fp32 (`bool`, *optional*, defaults to `False`):
|
| 79 |
+
Whether or not residuals should be in `float32`. If set to `False`
|
| 80 |
+
residuals will keep the same `dtype` as the rest of the model.
|
| 81 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 82 |
+
Whether or not the model should return the last key/values
|
| 83 |
+
attentions (not used by all models). Only relevant if
|
| 84 |
+
`config.is_decoder=True`.
|
| 85 |
+
num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
|
| 86 |
+
Number of prompt logits to calculate during generation. If `None`,
|
| 87 |
+
all logits will be calculated. If an integer value, only last
|
| 88 |
+
`num_logits_to_keep` logits will be calculated.
|
| 89 |
+
pad_token_id (`int`, *optional*, defaults to 0):
|
| 90 |
+
The id of the padding token.
|
| 91 |
+
bos_token_id (`int`, *optional*, defaults to 1):
|
| 92 |
+
The id of the "beginning-of-sequence" token.
|
| 93 |
+
eos_token_id (`int`, *optional*, defaults to 2):
|
| 94 |
+
The id of the "end-of-sequence" token.
|
| 95 |
+
sliding_window (`int`, *optional*, defaults to None):
|
| 96 |
+
Sliding window attention window size.
|
| 97 |
+
max_position_embeddings (`int`, *optional*, defaults to 4096):
|
| 98 |
+
The maximum sequence length that this model might ever be used
|
| 99 |
+
with.
|
| 100 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 101 |
+
The dropout ratio for the attention probabilities.
|
| 102 |
+
hidden_dropout (`float`, *optional*, defaults to 0.0):
|
| 103 |
+
The dropout ratio for the hidden states.
|
| 104 |
+
use_mamba_kernels (`bool`, *optional*, defaults to `True`):
|
| 105 |
+
Flag indicating whether or not to use the fast mamba kernels.
|
| 106 |
+
These are available only if `mamba-ssm` and `causal-conv1d`
|
| 107 |
+
are installed, and the mamba modules are running on a CUDA device.
|
| 108 |
+
ssm_state_size (`int`, *optional*, defaults to 128):
|
| 109 |
+
The dimension of the mamba state space latents.
|
| 110 |
+
mamba_num_heads (`int`, *optional*, defaults to 128):
|
| 111 |
+
Number of heads in Mamba layers.
|
| 112 |
+
mamba_n_groups (`int`, *optional*, defaults to 8):
|
| 113 |
+
Number of groups in Mamba layers.
|
| 114 |
+
mamba_head_dim (`int`, *optional*, defaults to 64):
|
| 115 |
+
Dimension of each Mamba head.
|
| 116 |
+
mamba_d_conv (`int`, *optional*, defaults to 4):
|
| 117 |
+
The size of the mamba convolution kernel.
|
| 118 |
+
mamba_expand (`int`, *optional*, defaults to 2):
|
| 119 |
+
Expanding factor used to determine the mamba intermediate size.
|
| 120 |
+
mamba_hidden_act (`str`, *optional*, defaults to "silu"):
|
| 121 |
+
The non-linear activation function in the Mamba layers.
|
| 122 |
+
mamba_dt_min (`float`, *optional*, defaults to 0.001):
|
| 123 |
+
Minimum value for the time step in Mamba.
|
| 124 |
+
mamba_dt_max (`float`, *optional*, defaults to 0.1):
|
| 125 |
+
Maximum value for the time step in Mamba.
|
| 126 |
+
mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
|
| 127 |
+
Limits for the time step in Mamba.
|
| 128 |
+
mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
|
| 129 |
+
Floor value for time step initialization in Mamba.
|
| 130 |
+
mamba_conv_bias (`bool`, *optional*, defaults to `True`):
|
| 131 |
+
Whether to use bias in the convolution layer of the mamba mixer
|
| 132 |
+
block.
|
| 133 |
+
mamba_proj_bias (`bool`, *optional*, defaults to `False`):
|
| 134 |
+
Whether to use bias in the input and output projections of the
|
| 135 |
+
mamba mixer block.
|
| 136 |
+
mamba_chunk_size (`int`, *optional*, defaults to 256):
|
| 137 |
+
Size of chunks for Mamba processing.
|
| 138 |
+
rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
|
| 139 |
+
Whether to rescale the pre-normalization residual connections.
|
| 140 |
+
"""
|
| 141 |
+
|
| 142 |
+
model_type = "nemotron_h"
|
| 143 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 144 |
+
|
| 145 |
+
def __init__(
|
| 146 |
+
self,
|
| 147 |
+
vocab_size=131072,
|
| 148 |
+
tie_word_embeddings=False,
|
| 149 |
+
hidden_size=4096,
|
| 150 |
+
intermediate_size=21504,
|
| 151 |
+
num_hidden_layers=52,
|
| 152 |
+
hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
|
| 153 |
+
num_attention_heads=32,
|
| 154 |
+
head_dim=128,
|
| 155 |
+
num_key_value_heads=8, # nemo: num_query_groups
|
| 156 |
+
mlp_hidden_act="relu2",
|
| 157 |
+
attention_bias=False,
|
| 158 |
+
mlp_bias=False,
|
| 159 |
+
use_bias=False,
|
| 160 |
+
initializer_range=0.02, # nemo: init_method_std
|
| 161 |
+
layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon
|
| 162 |
+
residual_in_fp32=False, # Megatron Core default value
|
| 163 |
+
use_cache=True,
|
| 164 |
+
num_logits_to_keep=1,
|
| 165 |
+
pad_token_id=0,
|
| 166 |
+
bos_token_id=1,
|
| 167 |
+
eos_token_id=2,
|
| 168 |
+
sliding_window=None,
|
| 169 |
+
max_position_embeddings=4096,
|
| 170 |
+
attention_dropout=0.0,
|
| 171 |
+
hidden_dropout=0.0, # * ADDED
|
| 172 |
+
use_mamba_kernels=True,
|
| 173 |
+
ssm_state_size=128, # mamba_state_size
|
| 174 |
+
mamba_num_heads=128,
|
| 175 |
+
mamba_n_groups=8, # nemo: mamba_ssm_ngroups = num_heads
|
| 176 |
+
mamba_head_dim=64,
|
| 177 |
+
mamba_d_conv=4,
|
| 178 |
+
mamba_expand=2,
|
| 179 |
+
mamba_hidden_act="silu",
|
| 180 |
+
mamba_dt_min=0.001,
|
| 181 |
+
mamba_dt_max=0.1,
|
| 182 |
+
mamba_dt_limit=(0.0, float("inf")),
|
| 183 |
+
mamba_dt_init_floor=1e-4,
|
| 184 |
+
mamba_conv_bias=True,
|
| 185 |
+
mamba_proj_bias=False,
|
| 186 |
+
mamba_chunk_size=256,
|
| 187 |
+
rescale_prenorm_residual=True,
|
| 188 |
+
**kwargs,
|
| 189 |
+
):
|
| 190 |
+
self.vocab_size = vocab_size
|
| 191 |
+
self.tie_word_embeddings = tie_word_embeddings
|
| 192 |
+
self.hidden_size = hidden_size
|
| 193 |
+
self.intermediate_size = intermediate_size
|
| 194 |
+
self.num_hidden_layers = num_hidden_layers
|
| 195 |
+
self.hybrid_override_pattern = hybrid_override_pattern
|
| 196 |
+
self.num_attention_heads = num_attention_heads
|
| 197 |
+
self.head_dim = head_dim
|
| 198 |
+
self.sliding_window = sliding_window
|
| 199 |
+
self.max_position_embeddings = max_position_embeddings
|
| 200 |
+
self.attention_dropout = attention_dropout
|
| 201 |
+
self.hidden_dropout = hidden_dropout
|
| 202 |
+
|
| 203 |
+
# Validate hybrid_override_pattern
|
| 204 |
+
# M: Mamba2, *: Attention, -: MLP
|
| 205 |
+
assert len(self.hybrid_override_pattern) == self.num_hidden_layers, (
|
| 206 |
+
"hybrid_override_pattern must have same length as "
|
| 207 |
+
"num_hidden_layers")
|
| 208 |
+
assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), (
|
| 209 |
+
"hybrid_override_pattern must only contain characters "
|
| 210 |
+
"'M', '*', or '-'")
|
| 211 |
+
|
| 212 |
+
# for backward compatibility
|
| 213 |
+
if num_key_value_heads is None:
|
| 214 |
+
num_key_value_heads = num_attention_heads
|
| 215 |
+
|
| 216 |
+
self.num_key_value_heads = num_key_value_heads
|
| 217 |
+
self.mlp_hidden_act = mlp_hidden_act
|
| 218 |
+
self.attention_bias = attention_bias
|
| 219 |
+
self.mlp_bias = mlp_bias
|
| 220 |
+
self.use_bias = use_bias
|
| 221 |
+
self.initializer_range = initializer_range
|
| 222 |
+
self.layer_norm_epsilon = layer_norm_epsilon
|
| 223 |
+
self.residual_in_fp32 = residual_in_fp32
|
| 224 |
+
|
| 225 |
+
self.use_cache = use_cache
|
| 226 |
+
self.num_logits_to_keep = num_logits_to_keep
|
| 227 |
+
|
| 228 |
+
self.use_mamba_kernels = use_mamba_kernels
|
| 229 |
+
self.n_groups = mamba_n_groups
|
| 230 |
+
self.mamba_head_dim = mamba_head_dim
|
| 231 |
+
self.ssm_state_size = ssm_state_size
|
| 232 |
+
self.mamba_num_heads = mamba_num_heads
|
| 233 |
+
self.conv_kernel = mamba_d_conv
|
| 234 |
+
self.expand = mamba_expand
|
| 235 |
+
self.mamba_hidden_act = mamba_hidden_act
|
| 236 |
+
self.time_step_min = mamba_dt_min
|
| 237 |
+
self.time_step_max = mamba_dt_max
|
| 238 |
+
self.time_step_limit = mamba_dt_limit
|
| 239 |
+
self.time_step_floor = mamba_dt_init_floor
|
| 240 |
+
self.use_conv_bias = mamba_conv_bias
|
| 241 |
+
self.mamba_proj_bias = mamba_proj_bias
|
| 242 |
+
self.chunk_size = mamba_chunk_size
|
| 243 |
+
self.rescale_prenorm_residual = rescale_prenorm_residual
|
| 244 |
+
|
| 245 |
+
super().__init__(
|
| 246 |
+
pad_token_id=pad_token_id,
|
| 247 |
+
bos_token_id=bos_token_id,
|
| 248 |
+
eos_token_id=eos_token_id,
|
| 249 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 250 |
+
**kwargs,
|
| 251 |
+
)
|
| 252 |
+
|
| 253 |
+
@property
|
| 254 |
+
def layers_block_type(self):
|
| 255 |
+
return [
|
| 256 |
+
"mamba" if self.hybrid_override_pattern[i] == "M" else
|
| 257 |
+
"attention" if self.hybrid_override_pattern[i] == "*" else "mlp"
|
| 258 |
+
for i in range(self.num_hidden_layers)
|
| 259 |
+
]
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/nemotron_vl.py
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# yapf: disable
|
| 5 |
+
# ruff: noqa: E501
|
| 6 |
+
# Adapted from
|
| 7 |
+
# https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/blob/main/configuration.py
|
| 8 |
+
# --------------------------------------------------------
|
| 9 |
+
# Adapted from https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B under MIT License
|
| 10 |
+
# LICENSE is in incl_licenses directory.
|
| 11 |
+
# --------------------------------------------------------
|
| 12 |
+
|
| 13 |
+
from transformers import LlamaConfig
|
| 14 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 15 |
+
from transformers.dynamic_module_utils import get_class_from_dynamic_module
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class Nemotron_Nano_VL_Config(PretrainedConfig):
|
| 19 |
+
model_type = 'Llama_Nemotron_Nano_VL'
|
| 20 |
+
is_composition = True
|
| 21 |
+
|
| 22 |
+
def __init__(
|
| 23 |
+
self,
|
| 24 |
+
vision_config=None,
|
| 25 |
+
llm_config=None,
|
| 26 |
+
force_image_size=None,
|
| 27 |
+
downsample_ratio=0.5,
|
| 28 |
+
template=None,
|
| 29 |
+
ps_version='v1',
|
| 30 |
+
image_tag_type="internvl",
|
| 31 |
+
projector_hidden_size=4096,
|
| 32 |
+
vit_hidden_size=1280,
|
| 33 |
+
**kwargs
|
| 34 |
+
):
|
| 35 |
+
super().__init__(**kwargs)
|
| 36 |
+
|
| 37 |
+
if vision_config is not None:
|
| 38 |
+
assert "auto_map" in vision_config and "AutoConfig" in vision_config["auto_map"]
|
| 39 |
+
vision_auto_config = get_class_from_dynamic_module(*vision_config["auto_map"]["AutoConfig"].split("--")[::-1])
|
| 40 |
+
self.vision_config = vision_auto_config(**vision_config)
|
| 41 |
+
else:
|
| 42 |
+
self.vision_config = PretrainedConfig()
|
| 43 |
+
|
| 44 |
+
if llm_config is None:
|
| 45 |
+
self.text_config = LlamaConfig()
|
| 46 |
+
else:
|
| 47 |
+
self.text_config = LlamaConfig(**llm_config)
|
| 48 |
+
|
| 49 |
+
# Assign configuration values
|
| 50 |
+
self.force_image_size = force_image_size
|
| 51 |
+
self.downsample_ratio = downsample_ratio
|
| 52 |
+
self.template = template # TODO move out of here and into the tokenizer
|
| 53 |
+
self.ps_version = ps_version # Pixel shuffle version
|
| 54 |
+
self.image_tag_type = image_tag_type # TODO: into the tokenizer too?
|
| 55 |
+
self.projector_hidden_size = projector_hidden_size
|
| 56 |
+
self.vit_hidden_size = vit_hidden_size
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/ovis.py
ADDED
|
@@ -0,0 +1,176 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# yapf: disable
|
| 5 |
+
# ruff: noqa: E501
|
| 6 |
+
# adapted from https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/configuration_aimv2.py
|
| 7 |
+
# and https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/configuration_ovis.py
|
| 8 |
+
# Ovis Config with AimV2 config registration removed for Transformers compatibility
|
| 9 |
+
from typing import Any, Optional, Union
|
| 10 |
+
|
| 11 |
+
from transformers import AutoConfig, PretrainedConfig
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
class AIMv2Config(PretrainedConfig):
|
| 15 |
+
"""This is the configuration class to store the configuration of an [`AIMv2Model`].
|
| 16 |
+
Instantiating a configuration with the defaults will yield a similar configuration
|
| 17 |
+
to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
|
| 18 |
+
Args:
|
| 19 |
+
hidden_size: Dimension of the hidden representations.
|
| 20 |
+
intermediate_size: Dimension of the SwiGLU representations.
|
| 21 |
+
num_hidden_layers: Number of hidden layers in the Transformer.
|
| 22 |
+
num_attention_heads: Number of attention heads for each attention layer
|
| 23 |
+
in the Transformer.
|
| 24 |
+
num_channels: Number of input channels.
|
| 25 |
+
image_size: Image size.
|
| 26 |
+
patch_size: Patch size.
|
| 27 |
+
rms_norm_eps: Epsilon value used for the RMS normalization layer.
|
| 28 |
+
attention_dropout: Dropout ratio for attention probabilities.
|
| 29 |
+
projection_dropout: Dropout ratio for the projection layer after the attention.
|
| 30 |
+
qkv_bias: Whether to add a bias to the queries, keys and values.
|
| 31 |
+
use_bias: Whether to add a bias in the feed-forward and projection layers.
|
| 32 |
+
kwargs: Keyword arguments for the [`PretrainedConfig`].
|
| 33 |
+
"""
|
| 34 |
+
|
| 35 |
+
model_type: str = "aimv2"
|
| 36 |
+
|
| 37 |
+
def __init__(
|
| 38 |
+
self,
|
| 39 |
+
hidden_size: int = 1024,
|
| 40 |
+
intermediate_size: int = 2816,
|
| 41 |
+
num_hidden_layers: int = 24,
|
| 42 |
+
num_attention_heads: int = 8,
|
| 43 |
+
num_channels: int = 3,
|
| 44 |
+
image_size: int = 224,
|
| 45 |
+
patch_size: int = 14,
|
| 46 |
+
rms_norm_eps: float = 1e-5,
|
| 47 |
+
attention_dropout: float = 0.0,
|
| 48 |
+
projection_dropout: float = 0.0,
|
| 49 |
+
qkv_bias: bool = False,
|
| 50 |
+
use_bias: bool = False,
|
| 51 |
+
**kwargs: Any,
|
| 52 |
+
):
|
| 53 |
+
super().__init__(**kwargs)
|
| 54 |
+
self.hidden_size = hidden_size
|
| 55 |
+
self.intermediate_size = intermediate_size
|
| 56 |
+
self.num_hidden_layers = num_hidden_layers
|
| 57 |
+
self.num_attention_heads = num_attention_heads
|
| 58 |
+
self.num_channels = num_channels
|
| 59 |
+
self.patch_size = patch_size
|
| 60 |
+
self.image_size = image_size
|
| 61 |
+
self.attention_dropout = attention_dropout
|
| 62 |
+
self.rms_norm_eps = rms_norm_eps
|
| 63 |
+
|
| 64 |
+
self.projection_dropout = projection_dropout
|
| 65 |
+
self.qkv_bias = qkv_bias
|
| 66 |
+
self.use_bias = use_bias
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
# ----------------------------------------------------------------------
|
| 70 |
+
# Visual Tokenizer Configuration
|
| 71 |
+
# ----------------------------------------------------------------------
|
| 72 |
+
class BaseVisualTokenizerConfig(PretrainedConfig):
|
| 73 |
+
|
| 74 |
+
def __init__(self,
|
| 75 |
+
vocab_size=16384,
|
| 76 |
+
tokenize_function="softmax",
|
| 77 |
+
tau=1.0,
|
| 78 |
+
depths=None,
|
| 79 |
+
drop_cls_token=False,
|
| 80 |
+
backbone_config: Optional[Union[PretrainedConfig,
|
| 81 |
+
dict]] = None,
|
| 82 |
+
hidden_stride: int = 1,
|
| 83 |
+
**kwargs):
|
| 84 |
+
super().__init__(**kwargs)
|
| 85 |
+
self.vocab_size = vocab_size
|
| 86 |
+
self.tokenize_function = tokenize_function
|
| 87 |
+
self.tau = tau
|
| 88 |
+
if isinstance(depths, str):
|
| 89 |
+
depths = [int(x) for x in depths.split('|')]
|
| 90 |
+
self.depths = depths
|
| 91 |
+
self.backbone_kwargs = dict[str, Any]()
|
| 92 |
+
self.drop_cls_token = drop_cls_token
|
| 93 |
+
if backbone_config is not None:
|
| 94 |
+
assert isinstance(backbone_config, (PretrainedConfig, dict)), \
|
| 95 |
+
f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"
|
| 96 |
+
if not isinstance(backbone_config, PretrainedConfig):
|
| 97 |
+
model_type = backbone_config['model_type']
|
| 98 |
+
if model_type != "aimv2":
|
| 99 |
+
backbone_config.pop('model_type')
|
| 100 |
+
backbone_config = AutoConfig.for_model(model_type, **backbone_config)
|
| 101 |
+
else:
|
| 102 |
+
backbone_config = AIMv2Config(**backbone_config)
|
| 103 |
+
self.backbone_config = backbone_config
|
| 104 |
+
self.hidden_stride = hidden_stride
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
class Aimv2VisualTokenizerConfig(BaseVisualTokenizerConfig):
|
| 108 |
+
model_type = "aimv2_visual_tokenizer"
|
| 109 |
+
|
| 110 |
+
def __init__(self, **kwargs):
|
| 111 |
+
super().__init__(**kwargs)
|
| 112 |
+
if self.drop_cls_token:
|
| 113 |
+
self.drop_cls_token = False
|
| 114 |
+
if self.depths:
|
| 115 |
+
assert len(self.depths) == 1
|
| 116 |
+
self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
class SiglipVisualTokenizerConfig(BaseVisualTokenizerConfig):
|
| 120 |
+
model_type = "siglip_visual_tokenizer"
|
| 121 |
+
|
| 122 |
+
def __init__(self, **kwargs):
|
| 123 |
+
super().__init__(**kwargs)
|
| 124 |
+
if self.drop_cls_token:
|
| 125 |
+
self.drop_cls_token = False
|
| 126 |
+
if self.depths:
|
| 127 |
+
assert len(self.depths) == 1
|
| 128 |
+
self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
AutoConfig.register("siglip_visual_tokenizer", SiglipVisualTokenizerConfig)
|
| 132 |
+
AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
# ----------------------------------------------------------------------
|
| 136 |
+
# Ovis Configuration
|
| 137 |
+
# ----------------------------------------------------------------------
|
| 138 |
+
class OvisConfig(PretrainedConfig):
|
| 139 |
+
model_type = "ovis"
|
| 140 |
+
|
| 141 |
+
def __init__(self,
|
| 142 |
+
llm_config: Optional[Union[PretrainedConfig, dict]] = None,
|
| 143 |
+
visual_tokenizer_config: Optional[Union[PretrainedConfig,
|
| 144 |
+
dict]] = None,
|
| 145 |
+
multimodal_max_length=8192,
|
| 146 |
+
hidden_size=None,
|
| 147 |
+
conversation_formatter_class=None,
|
| 148 |
+
llm_attn_implementation=None,
|
| 149 |
+
disable_tie_weight=False,
|
| 150 |
+
**kwargs):
|
| 151 |
+
super().__init__(**kwargs)
|
| 152 |
+
if llm_config is not None:
|
| 153 |
+
assert isinstance(llm_config, (PretrainedConfig, dict)), \
|
| 154 |
+
f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"
|
| 155 |
+
if not isinstance(llm_config, PretrainedConfig):
|
| 156 |
+
model_type = llm_config['model_type']
|
| 157 |
+
llm_config.pop('model_type')
|
| 158 |
+
llm_config = AutoConfig.for_model(model_type, **llm_config)
|
| 159 |
+
|
| 160 |
+
# map llm_config to text_config
|
| 161 |
+
self.text_config = llm_config
|
| 162 |
+
if visual_tokenizer_config is not None:
|
| 163 |
+
assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), \
|
| 164 |
+
f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
|
| 165 |
+
if not isinstance(visual_tokenizer_config, PretrainedConfig):
|
| 166 |
+
model_type = visual_tokenizer_config['model_type']
|
| 167 |
+
visual_tokenizer_config.pop('model_type')
|
| 168 |
+
visual_tokenizer_config = AutoConfig.for_model(
|
| 169 |
+
model_type, **visual_tokenizer_config)
|
| 170 |
+
|
| 171 |
+
self.visual_tokenizer_config = visual_tokenizer_config
|
| 172 |
+
self.multimodal_max_length = multimodal_max_length
|
| 173 |
+
self.hidden_size = hidden_size
|
| 174 |
+
self.conversation_formatter_class = conversation_formatter_class
|
| 175 |
+
self.llm_attn_implementation = llm_attn_implementation
|
| 176 |
+
self.disable_tie_weight = disable_tie_weight
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/step3_vl.py
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
from typing import Any, Optional, Union
|
| 4 |
+
|
| 5 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class Step3VisionEncoderConfig(PretrainedConfig):
|
| 9 |
+
model_type = "step3_vision_encoder"
|
| 10 |
+
|
| 11 |
+
def __init__(
|
| 12 |
+
self,
|
| 13 |
+
hidden_size=1792,
|
| 14 |
+
intermediate_size=3072,
|
| 15 |
+
output_hidden_size=4096,
|
| 16 |
+
num_hidden_layers=63,
|
| 17 |
+
num_attention_heads=16,
|
| 18 |
+
num_channels=3,
|
| 19 |
+
image_size=728,
|
| 20 |
+
patch_size=14,
|
| 21 |
+
hidden_act="quick_gelu",
|
| 22 |
+
layer_norm_eps=1e-5,
|
| 23 |
+
**kwargs,
|
| 24 |
+
):
|
| 25 |
+
self.hidden_size = hidden_size
|
| 26 |
+
self.intermediate_size = intermediate_size
|
| 27 |
+
self.output_hidden_size = output_hidden_size
|
| 28 |
+
self.num_hidden_layers = num_hidden_layers
|
| 29 |
+
self.num_attention_heads = num_attention_heads
|
| 30 |
+
self.num_channels = num_channels
|
| 31 |
+
self.patch_size = patch_size
|
| 32 |
+
self.image_size = image_size
|
| 33 |
+
self.layer_norm_eps = layer_norm_eps
|
| 34 |
+
self.hidden_act = hidden_act
|
| 35 |
+
super().__init__(**kwargs)
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
class Step3TextConfig(PretrainedConfig):
|
| 39 |
+
model_type = "step3_text"
|
| 40 |
+
architectures = ["Step3TextForCausalLM"]
|
| 41 |
+
|
| 42 |
+
def __init__(
|
| 43 |
+
self,
|
| 44 |
+
hidden_size: int = 7168,
|
| 45 |
+
intermediate_size: int = 18432,
|
| 46 |
+
num_attention_heads: int = 64,
|
| 47 |
+
num_attention_groups: int = 1,
|
| 48 |
+
num_hidden_layers: int = 61,
|
| 49 |
+
max_seq_len: int = 65536,
|
| 50 |
+
vocab_size: int = 128815,
|
| 51 |
+
rms_norm_eps: float = 1e-5,
|
| 52 |
+
moe_intermediate_size: int = 5120,
|
| 53 |
+
moe_num_experts: int = 48,
|
| 54 |
+
moe_top_k: int = 3,
|
| 55 |
+
rope_theta: float = 500000,
|
| 56 |
+
rope_scaling: Optional[dict[str, Any]] = None,
|
| 57 |
+
max_position_embedding: int = 65536,
|
| 58 |
+
share_expert_dim: int = 5120,
|
| 59 |
+
share_q_dim: int = 2048,
|
| 60 |
+
head_dim: int = 256,
|
| 61 |
+
norm_expert_weight: bool = False,
|
| 62 |
+
moe_layers_enum: tuple[int,
|
| 63 |
+
...] = (4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
|
| 64 |
+
15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
|
| 65 |
+
25, 26, 27, 28, 29, 30, 31, 32, 33, 34,
|
| 66 |
+
35, 36, 37, 38, 39, 40, 41, 42, 43, 44,
|
| 67 |
+
45, 46, 47, 48, 49, 50, 51, 52, 53, 54,
|
| 68 |
+
55, 56, 57, 58, 59),
|
| 69 |
+
**kwargs,
|
| 70 |
+
) -> None:
|
| 71 |
+
self.hidden_size = hidden_size
|
| 72 |
+
self.intermediate_size = intermediate_size
|
| 73 |
+
self.num_attention_heads = num_attention_heads
|
| 74 |
+
self.num_attention_groups = num_attention_groups
|
| 75 |
+
self.num_hidden_layers = num_hidden_layers
|
| 76 |
+
self.max_seq_len = max_seq_len
|
| 77 |
+
self.vocab_size = vocab_size
|
| 78 |
+
self.rms_norm_eps = rms_norm_eps
|
| 79 |
+
self.moe_intermediate_size = moe_intermediate_size
|
| 80 |
+
self.moe_num_experts = moe_num_experts
|
| 81 |
+
self.moe_top_k = moe_top_k
|
| 82 |
+
self.rope_theta = rope_theta
|
| 83 |
+
self.rope_scaling = rope_scaling
|
| 84 |
+
self.max_position_embedding = max_position_embedding
|
| 85 |
+
self.share_expert_dim = share_expert_dim
|
| 86 |
+
self.share_q_dim = share_q_dim
|
| 87 |
+
self.head_dim = head_dim
|
| 88 |
+
self.norm_expert_weight = norm_expert_weight
|
| 89 |
+
self.moe_layers_enum = moe_layers_enum
|
| 90 |
+
|
| 91 |
+
super().__init__(**kwargs)
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
class Step3VLConfig(PretrainedConfig):
|
| 95 |
+
model_type = "step3_vl"
|
| 96 |
+
|
| 97 |
+
def __init__(
|
| 98 |
+
self,
|
| 99 |
+
vision_config: Optional[Union[dict, Step3VisionEncoderConfig]] = None,
|
| 100 |
+
text_config: Optional[Union[dict, Step3TextConfig]] = None,
|
| 101 |
+
understand_projector_stride: int = 1,
|
| 102 |
+
projector_bias: bool = True,
|
| 103 |
+
image_token_id: int = 128001,
|
| 104 |
+
**kwargs,
|
| 105 |
+
) -> None:
|
| 106 |
+
if vision_config is None:
|
| 107 |
+
vision_config = Step3VisionEncoderConfig()
|
| 108 |
+
elif isinstance(vision_config, dict):
|
| 109 |
+
vision_config = Step3VisionEncoderConfig(**vision_config)
|
| 110 |
+
self.vision_config = vision_config
|
| 111 |
+
|
| 112 |
+
if text_config is None:
|
| 113 |
+
text_config = Step3TextConfig()
|
| 114 |
+
elif isinstance(text_config, dict):
|
| 115 |
+
text_config = Step3TextConfig(**text_config)
|
| 116 |
+
self.text_config = text_config
|
| 117 |
+
|
| 118 |
+
self.understand_projector_stride = understand_projector_stride
|
| 119 |
+
self.projector_bias = projector_bias
|
| 120 |
+
self.hidden_size = text_config.hidden_size
|
| 121 |
+
self.image_token_id = image_token_id
|
| 122 |
+
|
| 123 |
+
super().__init__(**kwargs)
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/configs/ultravox.py
ADDED
|
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# Adapted from https://github.com/fixie-ai/ultravox/blob/ecd58c4041030bae2ad15aa6bcf04ab43199ea02/ultravox/model/ultravox_config.py
|
| 5 |
+
from typing import Any, Optional
|
| 6 |
+
|
| 7 |
+
import transformers
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class UltravoxConfig(transformers.PretrainedConfig):
|
| 11 |
+
r"""
|
| 12 |
+
This is the configuration class to store the configuration of a
|
| 13 |
+
[`UltravoxForConditionalGeneration`]. It is used to instantiate an
|
| 14 |
+
Ultravox model according to the specified arguments, defining the model
|
| 15 |
+
architecture.
|
| 16 |
+
|
| 17 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to
|
| 18 |
+
control the model outputs. Read the documentation from [`PretrainedConfig`]
|
| 19 |
+
for more information.
|
| 20 |
+
|
| 21 |
+
Args:
|
| 22 |
+
audio_config (`Union[AutoConfig, dict]`, *optional*):
|
| 23 |
+
Custom audio config or dict
|
| 24 |
+
text_config (`Union[AutoConfig, dict]`, *optional*):
|
| 25 |
+
The config object of the text backbone. Can be any of `LlamaConfig`
|
| 26 |
+
or `MistralConfig`.
|
| 27 |
+
ignore_index (`int`, *optional*, defaults to -100):
|
| 28 |
+
The ignore index for the loss function.
|
| 29 |
+
audio_token_index (`int`, *optional*, defaults to 32000):
|
| 30 |
+
The audio token index to encode the audio prompt.
|
| 31 |
+
stack_factor (`int`, *optional*, defaults to 8):
|
| 32 |
+
Audio downsampling factor for the multimodal projector.
|
| 33 |
+
norm_init (`float`, *optional*, defaults to 0.4):
|
| 34 |
+
The initialization value for the layer normalization.
|
| 35 |
+
projector_act (`str`, *optional*, defaults to `"swiglu"`):
|
| 36 |
+
The activation function used by the multimodal projector.
|
| 37 |
+
text_model_lora_config (`LoraConfigSimplified`, *optional*):
|
| 38 |
+
The LoRA configuration for finetuning the text model.
|
| 39 |
+
audio_model_lora_config (`LoraConfigSimplified`, *optional*):
|
| 40 |
+
The LoRA configuration for finetuning the audio model.
|
| 41 |
+
projector_ln_mid (`bool`, *optional*, defaults to `False`):
|
| 42 |
+
Whether to apply layer normalization at the middle of the
|
| 43 |
+
projector or at the end. Versions v0.4.1 and below
|
| 44 |
+
use `False`, but v0.5 and above use `True`.
|
| 45 |
+
"""
|
| 46 |
+
|
| 47 |
+
model_type = "ultravox"
|
| 48 |
+
audio_token = "<|audio|>"
|
| 49 |
+
is_composition = False
|
| 50 |
+
|
| 51 |
+
def __init__(
|
| 52 |
+
self,
|
| 53 |
+
audio_config: Optional[dict[str, Any]] = None,
|
| 54 |
+
text_config: Optional[dict[str, Any]] = None,
|
| 55 |
+
audio_model_id: Optional[str] = None,
|
| 56 |
+
text_model_id: Optional[str] = None,
|
| 57 |
+
ignore_index: int = -100,
|
| 58 |
+
audio_token_index: int = 32000,
|
| 59 |
+
hidden_size: int = 4096,
|
| 60 |
+
stack_factor: int = 8,
|
| 61 |
+
norm_init: float = 0.4,
|
| 62 |
+
projector_act: str = "swiglu",
|
| 63 |
+
text_model_lora_config: Optional[dict[str, Any]] = None,
|
| 64 |
+
audio_model_lora_config: Optional[dict[str, Any]] = None,
|
| 65 |
+
projector_ln_mid: bool = False,
|
| 66 |
+
**kwargs,
|
| 67 |
+
):
|
| 68 |
+
self.ignore_index = ignore_index
|
| 69 |
+
|
| 70 |
+
self.audio_model_id = audio_model_id
|
| 71 |
+
self.text_model_id = text_model_id
|
| 72 |
+
self.audio_token_index = audio_token_index
|
| 73 |
+
|
| 74 |
+
self.hidden_size = hidden_size
|
| 75 |
+
self.stack_factor = stack_factor
|
| 76 |
+
self.norm_init = norm_init
|
| 77 |
+
self.projector_act = projector_act
|
| 78 |
+
self.projector_ln_mid = projector_ln_mid
|
| 79 |
+
|
| 80 |
+
if text_model_id is not None:
|
| 81 |
+
# Avoid circular import
|
| 82 |
+
from vllm.transformers_utils.config import get_config
|
| 83 |
+
|
| 84 |
+
text_config_obj = get_config(text_model_id,
|
| 85 |
+
trust_remote_code=False)
|
| 86 |
+
else:
|
| 87 |
+
text_config = text_config or {}
|
| 88 |
+
text_config_obj = transformers.CONFIG_MAPPING[text_config.get(
|
| 89 |
+
"model_type", "llama")](**text_config)
|
| 90 |
+
|
| 91 |
+
inner_text_config = text_config_obj.get_text_config()
|
| 92 |
+
|
| 93 |
+
if audio_model_id is not None:
|
| 94 |
+
# Avoid circular import
|
| 95 |
+
from vllm.transformers_utils.config import get_config
|
| 96 |
+
|
| 97 |
+
audio_config = get_config(audio_model_id, trust_remote_code=False)
|
| 98 |
+
else:
|
| 99 |
+
audio_config = audio_config or {}
|
| 100 |
+
audio_config = transformers.CONFIG_MAPPING[audio_config.get(
|
| 101 |
+
"model_type", "whisper")](**audio_config)
|
| 102 |
+
|
| 103 |
+
self.text_config = text_config_obj
|
| 104 |
+
self.audio_config = audio_config
|
| 105 |
+
self.text_model_lora_config = text_model_lora_config or {}
|
| 106 |
+
self.audio_model_lora_config = audio_model_lora_config or {}
|
| 107 |
+
|
| 108 |
+
self.vocab_size = inner_text_config.vocab_size
|
| 109 |
+
self.initializer_range = inner_text_config.initializer_range
|
| 110 |
+
self.text_hidden_size = inner_text_config.hidden_size
|
| 111 |
+
|
| 112 |
+
super().__init__(**kwargs)
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/__init__.py
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
"""
|
| 4 |
+
Multi-modal processors may be defined in this directory for the following
|
| 5 |
+
reasons:
|
| 6 |
+
|
| 7 |
+
- There is no processing file defined by HF Hub or Transformers library.
|
| 8 |
+
- There is a need to override the existing processor to support vLLM.
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from vllm.transformers_utils.processors.deepseek_vl2 import (
|
| 12 |
+
DeepseekVLV2Processor)
|
| 13 |
+
from vllm.transformers_utils.processors.ovis import OvisProcessor
|
| 14 |
+
|
| 15 |
+
__all__ = ["DeepseekVLV2Processor", "OvisProcessor"]
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/deepseek_vl2.py
ADDED
|
@@ -0,0 +1,363 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# yapf: disable
|
| 5 |
+
# ruff: noqa: E501
|
| 6 |
+
# coding=utf-8
|
| 7 |
+
# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/ff23960c5cf9e6874b44be38af930cfb0ccbb620/deepseek_vl2/models/processing_deepseek_vl_v2.py
|
| 8 |
+
# Copyright (c) 2023-2024 DeepSeek.
|
| 9 |
+
#
|
| 10 |
+
# Permission is hereby granted, free of charge, to any person obtaining a copy of
|
| 11 |
+
# this software and associated documentation files (the "Software"), to deal in
|
| 12 |
+
# the Software without restriction, including without limitation the rights to
|
| 13 |
+
# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
|
| 14 |
+
# the Software, and to permit persons to whom the Software is furnished to do so,
|
| 15 |
+
# subject to the following conditions:
|
| 16 |
+
#
|
| 17 |
+
# The above copyright notice and this permission notice shall be included in all
|
| 18 |
+
# copies or substantial portions of the Software.
|
| 19 |
+
#
|
| 20 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 21 |
+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
|
| 22 |
+
# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
|
| 23 |
+
# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
|
| 24 |
+
# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
|
| 25 |
+
# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
|
| 26 |
+
|
| 27 |
+
import math
|
| 28 |
+
|
| 29 |
+
import torch
|
| 30 |
+
import torchvision.transforms as T
|
| 31 |
+
from PIL import Image, ImageOps
|
| 32 |
+
from transformers import AutoProcessor, BatchFeature, LlamaTokenizerFast
|
| 33 |
+
from transformers.processing_utils import ProcessorMixin
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
class ImageTransform:
|
| 37 |
+
|
| 38 |
+
def __init__(self,
|
| 39 |
+
mean: tuple[float, float, float] = (0.5, 0.5, 0.5),
|
| 40 |
+
std: tuple[float, float, float] = (0.5, 0.5, 0.5),
|
| 41 |
+
normalize: bool = True):
|
| 42 |
+
self.mean = mean
|
| 43 |
+
self.std = std
|
| 44 |
+
self.normalize = normalize
|
| 45 |
+
|
| 46 |
+
transform_pipelines = [T.ToTensor()]
|
| 47 |
+
|
| 48 |
+
if normalize:
|
| 49 |
+
transform_pipelines.append(T.Normalize(mean, std))
|
| 50 |
+
|
| 51 |
+
self.transform = T.Compose(transform_pipelines)
|
| 52 |
+
|
| 53 |
+
def __call__(self, pil_img: Image.Image):
|
| 54 |
+
x = self.transform(pil_img)
|
| 55 |
+
return x
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
class DeepseekVLV2Processor(ProcessorMixin):
|
| 59 |
+
tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
|
| 60 |
+
attributes = ["tokenizer"]
|
| 61 |
+
|
| 62 |
+
def __init__(
|
| 63 |
+
self,
|
| 64 |
+
tokenizer: LlamaTokenizerFast,
|
| 65 |
+
candidate_resolutions: tuple[tuple[int, int]],
|
| 66 |
+
patch_size: int,
|
| 67 |
+
downsample_ratio: int,
|
| 68 |
+
image_mean: tuple[float, float, float] = (0.5, 0.5, 0.5),
|
| 69 |
+
image_std: tuple[float, float, float] = (0.5, 0.5, 0.5),
|
| 70 |
+
normalize: bool = True,
|
| 71 |
+
image_token: str = "<image>",
|
| 72 |
+
pad_token: str = "<|▁pad▁|>",
|
| 73 |
+
add_special_token: bool = False,
|
| 74 |
+
sft_format: str = "deepseek",
|
| 75 |
+
mask_prompt: bool = True,
|
| 76 |
+
ignore_id: int = -100,
|
| 77 |
+
**kwargs,
|
| 78 |
+
):
|
| 79 |
+
|
| 80 |
+
self.candidate_resolutions = candidate_resolutions
|
| 81 |
+
self.image_size = candidate_resolutions[0][0]
|
| 82 |
+
self.patch_size = patch_size
|
| 83 |
+
self.image_mean = image_mean
|
| 84 |
+
self.image_std = image_std
|
| 85 |
+
self.normalize = normalize
|
| 86 |
+
self.downsample_ratio = downsample_ratio
|
| 87 |
+
|
| 88 |
+
self.image_transform = ImageTransform(mean=image_mean, std=image_std, normalize=normalize)
|
| 89 |
+
self.tokenizer = tokenizer
|
| 90 |
+
self.tokenizer.padding_side = 'left' # must set this,padding side with make a difference in batch inference
|
| 91 |
+
|
| 92 |
+
# add the pad_token as special token to use 'tokenizer.pad_token' and 'tokenizer.pad_token_id'
|
| 93 |
+
if tokenizer.pad_token is None:
|
| 94 |
+
self.tokenizer.add_special_tokens({'pad_token': pad_token})
|
| 95 |
+
|
| 96 |
+
# add image token
|
| 97 |
+
image_token_id = self.tokenizer.vocab.get(image_token)
|
| 98 |
+
if image_token_id is None:
|
| 99 |
+
special_tokens = [image_token]
|
| 100 |
+
special_tokens_dict = {"additional_special_tokens": special_tokens}
|
| 101 |
+
self.tokenizer.add_special_tokens(special_tokens_dict)
|
| 102 |
+
self.image_token_id = self.tokenizer.vocab.get(image_token)
|
| 103 |
+
|
| 104 |
+
# add five special tokens for grounding-related tasks
|
| 105 |
+
# <|ref|>, <|/ref|>, <|det|>, <|/det|>, <|grounding|>
|
| 106 |
+
special_tokens = ['<|ref|>', '<|/ref|>', '<|det|>', '<|/det|>', '<|grounding|>']
|
| 107 |
+
special_tokens_dict = {"additional_special_tokens": special_tokens}
|
| 108 |
+
self.tokenizer.add_special_tokens(special_tokens_dict)
|
| 109 |
+
|
| 110 |
+
# add special tokens for SFT data
|
| 111 |
+
special_tokens = ["<|User|>", "<|Assistant|>"]
|
| 112 |
+
special_tokens_dict = {"additional_special_tokens": special_tokens}
|
| 113 |
+
self.tokenizer.add_special_tokens(special_tokens_dict)
|
| 114 |
+
|
| 115 |
+
self.image_token = image_token
|
| 116 |
+
self.pad_token = pad_token
|
| 117 |
+
self.add_special_token = add_special_token
|
| 118 |
+
self.sft_format = sft_format
|
| 119 |
+
self.mask_prompt = mask_prompt
|
| 120 |
+
self.ignore_id = ignore_id
|
| 121 |
+
|
| 122 |
+
super().__init__(
|
| 123 |
+
tokenizer,
|
| 124 |
+
**kwargs,
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
def select_best_resolution(self, image_size):
|
| 128 |
+
# used for cropping
|
| 129 |
+
original_width, original_height = image_size
|
| 130 |
+
best_fit = None
|
| 131 |
+
max_effective_resolution = 0
|
| 132 |
+
min_wasted_resolution = float("inf")
|
| 133 |
+
|
| 134 |
+
for width, height in self.candidate_resolutions:
|
| 135 |
+
scale = min(width / original_width, height / original_height)
|
| 136 |
+
downscaled_width, downscaled_height = int(
|
| 137 |
+
original_width * scale), int(original_height * scale)
|
| 138 |
+
effective_resolution = min(downscaled_width * downscaled_height,
|
| 139 |
+
original_width * original_height)
|
| 140 |
+
wasted_resolution = (width * height) - effective_resolution
|
| 141 |
+
|
| 142 |
+
if effective_resolution > max_effective_resolution or (
|
| 143 |
+
effective_resolution == max_effective_resolution
|
| 144 |
+
and wasted_resolution < min_wasted_resolution):
|
| 145 |
+
max_effective_resolution = effective_resolution
|
| 146 |
+
min_wasted_resolution = wasted_resolution
|
| 147 |
+
best_fit = (width, height)
|
| 148 |
+
|
| 149 |
+
return best_fit
|
| 150 |
+
|
| 151 |
+
@property
|
| 152 |
+
def bos_id(self):
|
| 153 |
+
return self.tokenizer.bos_token_id
|
| 154 |
+
|
| 155 |
+
@property
|
| 156 |
+
def eos_id(self):
|
| 157 |
+
return self.tokenizer.eos_token_id
|
| 158 |
+
|
| 159 |
+
@property
|
| 160 |
+
def pad_id(self):
|
| 161 |
+
return self.tokenizer.pad_token_id
|
| 162 |
+
|
| 163 |
+
def encode(self, text: str, bos: bool = True, eos: bool = False):
|
| 164 |
+
t = self.tokenizer.encode(text, add_special_tokens=False)
|
| 165 |
+
|
| 166 |
+
if bos:
|
| 167 |
+
t = [self.bos_id] + t
|
| 168 |
+
if eos:
|
| 169 |
+
t = t + [self.eos_id]
|
| 170 |
+
|
| 171 |
+
return t
|
| 172 |
+
|
| 173 |
+
def decode(self, t: list[int], **kwargs) -> str:
|
| 174 |
+
return self.tokenizer.decode(t, **kwargs)
|
| 175 |
+
|
| 176 |
+
def process_one(
|
| 177 |
+
self,
|
| 178 |
+
prompt: str,
|
| 179 |
+
images: list[Image.Image],
|
| 180 |
+
inference_mode: bool = True,
|
| 181 |
+
**kwargs,
|
| 182 |
+
):
|
| 183 |
+
"""
|
| 184 |
+
|
| 185 |
+
Args:
|
| 186 |
+
prompt (str): the formatted prompt;
|
| 187 |
+
conversations (list[dict]): conversations with a list of messages;
|
| 188 |
+
images (list[ImageType]): the list of images;
|
| 189 |
+
inference_mode (bool): if True, then remove the last eos token;
|
| 190 |
+
system_prompt (str): the system prompt;
|
| 191 |
+
**kwargs:
|
| 192 |
+
|
| 193 |
+
Returns:
|
| 194 |
+
outputs (BaseProcessorOutput): the output of the processor,
|
| 195 |
+
- input_ids (torch.LongTensor): [N + image tokens]
|
| 196 |
+
- target_ids (torch.LongTensor): [N + image tokens]
|
| 197 |
+
- pixel_values (torch.FloatTensor): [n_patches, 3, H, W]
|
| 198 |
+
- image_id (int): the id of the image token
|
| 199 |
+
- num_image_tokens (list[int]): the number of image tokens
|
| 200 |
+
"""
|
| 201 |
+
|
| 202 |
+
assert (prompt is not None and images is not None
|
| 203 |
+
), "prompt and images must be used at the same time."
|
| 204 |
+
|
| 205 |
+
sft_format = prompt
|
| 206 |
+
tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens = self.tokenize_with_images(
|
| 207 |
+
sft_format, images, bos=True, eos=True, cropping=len(images) <= 2)
|
| 208 |
+
masked_tokenized_str = []
|
| 209 |
+
for token_index in tokenized_str:
|
| 210 |
+
if token_index != self.image_token_id:
|
| 211 |
+
masked_tokenized_str.append(token_index)
|
| 212 |
+
else:
|
| 213 |
+
masked_tokenized_str.append(self.ignore_id)
|
| 214 |
+
|
| 215 |
+
assert len(tokenized_str) == len(images_seq_mask) == len(masked_tokenized_str), \
|
| 216 |
+
(f"tokenized_str's length {len(tokenized_str)}, input_ids' length {len(masked_tokenized_str)}, "
|
| 217 |
+
f"imags_seq_mask's length {len(images_seq_mask)}, are not equal")
|
| 218 |
+
|
| 219 |
+
input_ids = torch.LongTensor(tokenized_str)
|
| 220 |
+
target_ids = torch.LongTensor(masked_tokenized_str)
|
| 221 |
+
images_seq_mask = torch.tensor(images_seq_mask, dtype=torch.bool)
|
| 222 |
+
|
| 223 |
+
# set input_ids < 0 | input_ids == self.image_token_id as ignore_id
|
| 224 |
+
target_ids[(input_ids < 0) |
|
| 225 |
+
(input_ids == self.image_token_id)] = self.ignore_id
|
| 226 |
+
input_ids[input_ids < 0] = self.pad_id
|
| 227 |
+
|
| 228 |
+
if inference_mode:
|
| 229 |
+
# Remove the ending eos token
|
| 230 |
+
assert input_ids[-1] == self.eos_id
|
| 231 |
+
input_ids = input_ids[:-1]
|
| 232 |
+
target_ids = target_ids[:-1]
|
| 233 |
+
images_seq_mask = images_seq_mask[:-1]
|
| 234 |
+
|
| 235 |
+
if len(images_list) == 0:
|
| 236 |
+
pixel_values = torch.zeros((1, 3, self.image_size, self.image_size))
|
| 237 |
+
images_spatial_crop = torch.zeros((1, 2), dtype=torch.long)
|
| 238 |
+
else:
|
| 239 |
+
pixel_values = torch.stack(images_list, dim=0)
|
| 240 |
+
images_spatial_crop = torch.tensor(images_spatial_crop, dtype=torch.long)
|
| 241 |
+
|
| 242 |
+
input_ids = input_ids.unsqueeze(0)
|
| 243 |
+
|
| 244 |
+
prepare = BatchFeature(
|
| 245 |
+
data=dict(
|
| 246 |
+
input_ids=input_ids,
|
| 247 |
+
pixel_values=pixel_values,
|
| 248 |
+
images_seq_mask=images_seq_mask,
|
| 249 |
+
images_spatial_crop=images_spatial_crop,
|
| 250 |
+
num_image_tokens=num_image_tokens,
|
| 251 |
+
),
|
| 252 |
+
tensor_type="pt",
|
| 253 |
+
)
|
| 254 |
+
return prepare
|
| 255 |
+
|
| 256 |
+
def __call__(
|
| 257 |
+
self,
|
| 258 |
+
*,
|
| 259 |
+
text: str,
|
| 260 |
+
images: list[Image.Image],
|
| 261 |
+
inference_mode: bool = True,
|
| 262 |
+
**kwargs,
|
| 263 |
+
):
|
| 264 |
+
"""
|
| 265 |
+
|
| 266 |
+
Args:
|
| 267 |
+
text (str): the formatted prompt;
|
| 268 |
+
images (list[ImageType]): the list of images;
|
| 269 |
+
inference_mode (bool): if True, then remove the last eos token;
|
| 270 |
+
**kwargs:
|
| 271 |
+
|
| 272 |
+
Returns:
|
| 273 |
+
outputs (BaseProcessorOutput): the output of the processor,
|
| 274 |
+
- input_ids (torch.LongTensor): [N + image tokens]
|
| 275 |
+
- images (torch.FloatTensor): [n_images, 3, H, W]
|
| 276 |
+
- image_id (int): the id of the image token
|
| 277 |
+
- num_image_tokens (list[int]): the number of image tokens
|
| 278 |
+
"""
|
| 279 |
+
|
| 280 |
+
prepare = self.process_one(
|
| 281 |
+
prompt=text,
|
| 282 |
+
images=images,
|
| 283 |
+
inference_mode=inference_mode,
|
| 284 |
+
)
|
| 285 |
+
|
| 286 |
+
return prepare
|
| 287 |
+
|
| 288 |
+
def tokenize_with_images(
|
| 289 |
+
self,
|
| 290 |
+
conversation: str,
|
| 291 |
+
images: list[Image.Image],
|
| 292 |
+
bos: bool = True,
|
| 293 |
+
eos: bool = True,
|
| 294 |
+
cropping: bool = True,
|
| 295 |
+
):
|
| 296 |
+
"""Tokenize text with <image> tags."""
|
| 297 |
+
assert conversation.count(self.image_token) == len(images)
|
| 298 |
+
text_splits = conversation.split(self.image_token)
|
| 299 |
+
images_list, images_seq_mask, images_spatial_crop = [], [], []
|
| 300 |
+
num_image_tokens = []
|
| 301 |
+
tokenized_str = []
|
| 302 |
+
for text_sep, image in zip(text_splits, images):
|
| 303 |
+
"""encode text_sep"""
|
| 304 |
+
tokenized_sep = self.encode(text_sep, bos=False, eos=False)
|
| 305 |
+
tokenized_str += tokenized_sep
|
| 306 |
+
images_seq_mask += [False] * len(tokenized_sep)
|
| 307 |
+
|
| 308 |
+
"""select best resolution for anyres"""
|
| 309 |
+
if cropping:
|
| 310 |
+
best_width, best_height = self.select_best_resolution(image.size)
|
| 311 |
+
else:
|
| 312 |
+
best_width, best_height = self.image_size, self.image_size
|
| 313 |
+
|
| 314 |
+
"""process the global view"""
|
| 315 |
+
global_view = ImageOps.pad(image, (self.image_size, self.image_size),
|
| 316 |
+
color=tuple(int(x * 255) for x in self.image_transform.mean))
|
| 317 |
+
images_list.append(self.image_transform(global_view))
|
| 318 |
+
|
| 319 |
+
"""process the local views"""
|
| 320 |
+
local_view = ImageOps.pad(image, (best_width, best_height),
|
| 321 |
+
color=tuple(int(x * 255) for x in self.image_transform.mean))
|
| 322 |
+
for i in range(0, best_height, self.image_size):
|
| 323 |
+
for j in range(0, best_width, self.image_size):
|
| 324 |
+
images_list.append(
|
| 325 |
+
self.image_transform(local_view.crop((j, i, j + self.image_size, i + self.image_size))))
|
| 326 |
+
|
| 327 |
+
"""record height / width crop num"""
|
| 328 |
+
num_width_tiles, num_height_tiles = best_width // self.image_size, best_height // self.image_size
|
| 329 |
+
images_spatial_crop.append([num_width_tiles, num_height_tiles])
|
| 330 |
+
|
| 331 |
+
"""add image tokens"""
|
| 332 |
+
h = w = math.ceil((self.image_size // self.patch_size) / self.downsample_ratio)
|
| 333 |
+
# global views tokens h * (w + 1), 1 is for line separator
|
| 334 |
+
tokenized_image = [self.image_token_id] * h * (w + 1)
|
| 335 |
+
# add a separator between global and local views
|
| 336 |
+
tokenized_image += [self.image_token_id]
|
| 337 |
+
# local views tokens, (num_height_tiles * h) * (num_width_tiles * w + 1)
|
| 338 |
+
tokenized_image += [self.image_token_id] * (num_height_tiles * h) * (num_width_tiles * w + 1)
|
| 339 |
+
|
| 340 |
+
tokenized_str += tokenized_image
|
| 341 |
+
images_seq_mask += [True] * len(tokenized_image)
|
| 342 |
+
num_image_tokens.append(len(tokenized_image))
|
| 343 |
+
|
| 344 |
+
"""process the last text split"""
|
| 345 |
+
tokenized_sep = self.encode(text_splits[-1], bos=False, eos=False)
|
| 346 |
+
tokenized_str += tokenized_sep
|
| 347 |
+
images_seq_mask += [False] * len(tokenized_sep)
|
| 348 |
+
|
| 349 |
+
"""add the bos and eos tokens"""
|
| 350 |
+
if bos:
|
| 351 |
+
tokenized_str = [self.bos_id] + tokenized_str
|
| 352 |
+
images_seq_mask = [False] + images_seq_mask
|
| 353 |
+
if eos:
|
| 354 |
+
tokenized_str = tokenized_str + [self.eos_id]
|
| 355 |
+
images_seq_mask = images_seq_mask + [False]
|
| 356 |
+
|
| 357 |
+
assert len(tokenized_str) == len(
|
| 358 |
+
images_seq_mask), f"tokenize_with_images func: tokenized_str's length {len(tokenized_str)} is not equal to imags_seq_mask's length {len(images_seq_mask)}"
|
| 359 |
+
|
| 360 |
+
return tokenized_str, images_list, images_seq_mask, images_spatial_crop, num_image_tokens
|
| 361 |
+
|
| 362 |
+
|
| 363 |
+
AutoProcessor.register("DeepseekVLV2Processor", DeepseekVLV2Processor)
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/processors/ovis.py
ADDED
|
@@ -0,0 +1,420 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
# yapf: disable
|
| 5 |
+
# ruff: noqa: E501
|
| 6 |
+
# coding=utf-8
|
| 7 |
+
# adapted from https://github.com/AIDC-AI/Ovis/blob/35ab51a1a1e3542fa6db260a1084cefbc8f164bb/ovis/vllm/processing_ovis.py
|
| 8 |
+
# Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
|
| 9 |
+
#
|
| 10 |
+
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
| 11 |
+
# and OPT implementations in this library. It has been modified from its
|
| 12 |
+
# original forms to accommodate minor architectural differences compared
|
| 13 |
+
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
|
| 14 |
+
#
|
| 15 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 16 |
+
# you may not use this file except in compliance with the License.
|
| 17 |
+
# You may obtain a copy of the License at
|
| 18 |
+
#
|
| 19 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 20 |
+
#
|
| 21 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 22 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 23 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 24 |
+
# See the License for the specific language governing permissions and
|
| 25 |
+
# limitations under the License.
|
| 26 |
+
from functools import cached_property
|
| 27 |
+
from typing import Union
|
| 28 |
+
|
| 29 |
+
import PIL
|
| 30 |
+
import torch
|
| 31 |
+
from transformers import AutoProcessor, BatchFeature
|
| 32 |
+
from transformers.image_utils import ImageInput
|
| 33 |
+
from transformers.processing_utils import (ProcessingKwargs, ProcessorMixin,
|
| 34 |
+
Unpack)
|
| 35 |
+
from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
|
| 36 |
+
|
| 37 |
+
from vllm.multimodal.image import convert_image_mode
|
| 38 |
+
|
| 39 |
+
__all__ = ['OvisProcessor']
|
| 40 |
+
IGNORE_ID = -100
|
| 41 |
+
|
| 42 |
+
class OvisProcessorKwargs(ProcessingKwargs, total=False): # type: ignore[call-arg]
|
| 43 |
+
_defaults = {
|
| 44 |
+
"text_kwargs": {
|
| 45 |
+
"padding": False,
|
| 46 |
+
},
|
| 47 |
+
"images_kwargs": {
|
| 48 |
+
'max_partition':9,
|
| 49 |
+
'covering_threshold':0.9,
|
| 50 |
+
'convert_to_rgb':True,
|
| 51 |
+
'return_tensors':'pt'},
|
| 52 |
+
}
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
class OvisProcessor(ProcessorMixin):
|
| 57 |
+
r"""
|
| 58 |
+
Constructs a Ovis processor which wraps a Ovis image processor and a Qwen2 tokenizer into a single processor.
|
| 59 |
+
[`OvisProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
|
| 60 |
+
[`~OvisProcessor.__call__`] and [`~OvisProcessor.decode`] for more information.
|
| 61 |
+
Args:
|
| 62 |
+
image_processor ([`Qwen2VLImageProcessor`], *optional*):
|
| 63 |
+
The image processor is a required input.
|
| 64 |
+
tokenizer ([`Qwen2TokenizerFast`], *optional*):
|
| 65 |
+
The tokenizer is a required input.
|
| 66 |
+
chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
|
| 67 |
+
in a chat into a tokenizable string.
|
| 68 |
+
"""
|
| 69 |
+
|
| 70 |
+
attributes = ["image_processor", "tokenizer"]
|
| 71 |
+
valid_kwargs = ["chat_template", "image_pad_token", "image_segment_len"]
|
| 72 |
+
|
| 73 |
+
image_processor_class = "AutoImageProcessor"
|
| 74 |
+
tokenizer_class = "AutoTokenizer"
|
| 75 |
+
|
| 76 |
+
def __init__(
|
| 77 |
+
self,
|
| 78 |
+
image_processor=None,
|
| 79 |
+
tokenizer=None,
|
| 80 |
+
chat_template=None,
|
| 81 |
+
image_pad_token=None,
|
| 82 |
+
image_segment_len=255,
|
| 83 |
+
**kwargs,
|
| 84 |
+
):
|
| 85 |
+
self.image_token = "<image>"
|
| 86 |
+
self.image_pad_token = image_pad_token
|
| 87 |
+
self.image_segment_len = image_segment_len
|
| 88 |
+
super().__init__(image_processor, tokenizer, chat_template=chat_template)
|
| 89 |
+
|
| 90 |
+
@cached_property
|
| 91 |
+
def extra_special_tokens(self):
|
| 92 |
+
image_pad_token_id = self.tokenizer.get_vocab()[self.image_pad_token]
|
| 93 |
+
extra_special_tokens = {
|
| 94 |
+
"image_token": -200,
|
| 95 |
+
"image_atom": -300,
|
| 96 |
+
"image_start": -301,
|
| 97 |
+
"image_prefix": -302,
|
| 98 |
+
"image_col_sep": -303,
|
| 99 |
+
"image_row_sep": -304,
|
| 100 |
+
"image_end": -305,
|
| 101 |
+
'image_pad': image_pad_token_id,
|
| 102 |
+
}
|
| 103 |
+
return extra_special_tokens
|
| 104 |
+
|
| 105 |
+
def __call__(
|
| 106 |
+
self,
|
| 107 |
+
images: ImageInput = None,
|
| 108 |
+
text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
|
| 109 |
+
**kwargs: Unpack[OvisProcessorKwargs],
|
| 110 |
+
) -> BatchFeature:
|
| 111 |
+
"""
|
| 112 |
+
Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
|
| 113 |
+
and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
|
| 114 |
+
the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
|
| 115 |
+
Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
|
| 116 |
+
Args:
|
| 117 |
+
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
|
| 118 |
+
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
|
| 119 |
+
tensor. Both channels-first and channels-last formats are supported.
|
| 120 |
+
text (`str`, `list[str]`, `list[list[str]]`):
|
| 121 |
+
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
|
| 122 |
+
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
|
| 123 |
+
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
|
| 124 |
+
videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
|
| 125 |
+
The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
|
| 126 |
+
tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
|
| 127 |
+
return_tensors (`str` or [`~utils.TensorType`], *optional*):
|
| 128 |
+
If set, will return tensors of a particular framework. Acceptable values are:
|
| 129 |
+
- `'tf'`: Return TensorFlow `tf.constant` objects.
|
| 130 |
+
- `'pt'`: Return PyTorch `torch.Tensor` objects.
|
| 131 |
+
- `'np'`: Return NumPy `np.ndarray` objects.
|
| 132 |
+
- `'jax'`: Return JAX `jnp.ndarray` objects.
|
| 133 |
+
Returns:
|
| 134 |
+
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
|
| 135 |
+
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
|
| 136 |
+
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
|
| 137 |
+
`return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
|
| 138 |
+
`None`).
|
| 139 |
+
- **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
|
| 140 |
+
- **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.
|
| 141 |
+
- **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
|
| 142 |
+
- **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.
|
| 143 |
+
- **second_per_grid_ts** -- List of video seconds per time grid. Returned when `videos` is not `None`.
|
| 144 |
+
"""
|
| 145 |
+
output_kwargs = self._merge_kwargs(
|
| 146 |
+
OvisProcessorKwargs,
|
| 147 |
+
tokenizer_init_kwargs=self.tokenizer.init_kwargs,
|
| 148 |
+
**kwargs,
|
| 149 |
+
)
|
| 150 |
+
|
| 151 |
+
# Process all images first
|
| 152 |
+
image_features = {}
|
| 153 |
+
if images is not None:
|
| 154 |
+
processed_images = []
|
| 155 |
+
image_placeholders_list = []
|
| 156 |
+
grids = []
|
| 157 |
+
|
| 158 |
+
# Process each image
|
| 159 |
+
for image in images if isinstance(images, list) else [images]:
|
| 160 |
+
pixel_values, image_placeholders, grid = self.preprocess_image(
|
| 161 |
+
image=image, **output_kwargs["images_kwargs"]
|
| 162 |
+
)
|
| 163 |
+
processed_images.append(pixel_values)
|
| 164 |
+
image_placeholders_list.append(image_placeholders)
|
| 165 |
+
grids.append(grid)
|
| 166 |
+
|
| 167 |
+
# assign all processed images
|
| 168 |
+
if processed_images:
|
| 169 |
+
image_features["image_placeholders"] = image_placeholders_list
|
| 170 |
+
|
| 171 |
+
# Process text input
|
| 172 |
+
if text is not None:
|
| 173 |
+
|
| 174 |
+
if not isinstance(text, list):
|
| 175 |
+
text = [text]
|
| 176 |
+
|
| 177 |
+
tokenized_batched_text = self._tokenize_with_image_symbol(text)
|
| 178 |
+
image_token_id = self.get_token_value("image_token")
|
| 179 |
+
replaced_ids_list = []
|
| 180 |
+
idx = 0
|
| 181 |
+
for ids_tensor in tokenized_batched_text:
|
| 182 |
+
if image_token_id in ids_tensor and "image_placeholders" in image_features:
|
| 183 |
+
if idx < len(image_features["image_placeholders"]):
|
| 184 |
+
# Converts in list for ease of use
|
| 185 |
+
ids_list = ids_tensor.tolist()
|
| 186 |
+
|
| 187 |
+
new_ids = []
|
| 188 |
+
|
| 189 |
+
# replace placeholders
|
| 190 |
+
for i, token_id in enumerate(ids_list):
|
| 191 |
+
if token_id == image_token_id:
|
| 192 |
+
placeholder_ids = image_features["image_placeholders"][idx]
|
| 193 |
+
new_ids.extend(placeholder_ids)
|
| 194 |
+
idx += 1
|
| 195 |
+
else:
|
| 196 |
+
new_ids.append(token_id)
|
| 197 |
+
|
| 198 |
+
# Converts back to tensors
|
| 199 |
+
ids_tensor = torch.tensor(new_ids, dtype=torch.long)
|
| 200 |
+
else:
|
| 201 |
+
raise RuntimeError(
|
| 202 |
+
'Mismatch between the images you provided and the number of placeholder present in the text')
|
| 203 |
+
|
| 204 |
+
replaced_ids_list.append(ids_tensor)
|
| 205 |
+
|
| 206 |
+
if replaced_ids_list:
|
| 207 |
+
replaced_and_tokenized_ids = torch.stack(replaced_ids_list)
|
| 208 |
+
else:
|
| 209 |
+
replaced_and_tokenized_ids = torch.tensor([], dtype=torch.long)
|
| 210 |
+
|
| 211 |
+
# Create the output with text features
|
| 212 |
+
output = BatchFeature(
|
| 213 |
+
data={
|
| 214 |
+
"input_ids": replaced_and_tokenized_ids,
|
| 215 |
+
}
|
| 216 |
+
)
|
| 217 |
+
|
| 218 |
+
# Add image features if present
|
| 219 |
+
if image_features:
|
| 220 |
+
output["pixel_values"] = processed_images
|
| 221 |
+
output['grids'] = grids
|
| 222 |
+
|
| 223 |
+
return output
|
| 224 |
+
|
| 225 |
+
# If only images were provided
|
| 226 |
+
return BatchFeature(data=image_features)
|
| 227 |
+
|
| 228 |
+
def _tokenize_with_image_symbol(self, text_list: list[str]) -> torch.LongTensor:
|
| 229 |
+
batch_token_ids = []
|
| 230 |
+
for text in text_list:
|
| 231 |
+
text_chunks = [self.tokenizer(chunk, add_special_tokens=False).input_ids for chunk in
|
| 232 |
+
text.split(self.image_token)]
|
| 233 |
+
token_ids = []
|
| 234 |
+
num_chuck = len(text_chunks)
|
| 235 |
+
for i, chunk in enumerate(text_chunks):
|
| 236 |
+
token_ids.extend(chunk)
|
| 237 |
+
if i < num_chuck - 1:
|
| 238 |
+
token_ids.append(self.get_token_value("image_token"))
|
| 239 |
+
batch_token_ids.append(token_ids)
|
| 240 |
+
return torch.tensor(batch_token_ids, dtype=torch.long)
|
| 241 |
+
|
| 242 |
+
def get_image_size(self):
|
| 243 |
+
size = self.image_processor.size
|
| 244 |
+
if 'shortest_edge' in size:
|
| 245 |
+
width = height = size['shortest_edge']
|
| 246 |
+
elif "height" in size and "width" in size:
|
| 247 |
+
width = size['width']
|
| 248 |
+
height = size['height']
|
| 249 |
+
else:
|
| 250 |
+
raise ValueError( "Can't parse image size from image_processor config.")
|
| 251 |
+
return height, width
|
| 252 |
+
|
| 253 |
+
def get_token_value(self, tok):
|
| 254 |
+
return self.extra_special_tokens[tok]
|
| 255 |
+
|
| 256 |
+
def construct_image_indicators(self, grid):
|
| 257 |
+
image_placeholders = [self.get_token_value('image_start'),
|
| 258 |
+
self.get_token_value('image_atom'),
|
| 259 |
+
self.get_token_value('image_prefix')]
|
| 260 |
+
if grid[0] * grid[1] > 1:
|
| 261 |
+
for r in range(grid[0]):
|
| 262 |
+
for c in range(grid[1]):
|
| 263 |
+
image_placeholders.append(self.get_token_value('image_atom') )
|
| 264 |
+
if c < grid[1] - 1:
|
| 265 |
+
image_placeholders.append(self.get_token_value('image_col_sep'))
|
| 266 |
+
if r < grid[0] - 1:
|
| 267 |
+
image_placeholders.append(self.get_token_value('image_row_sep'))
|
| 268 |
+
image_placeholders.append(self.get_token_value('image_end'))
|
| 269 |
+
return image_placeholders
|
| 270 |
+
|
| 271 |
+
def construct_image_placeholders(self, grid):
|
| 272 |
+
|
| 273 |
+
image_placeholders = self.construct_image_indicators(grid)
|
| 274 |
+
|
| 275 |
+
image_atom_token_id = self.get_token_value('image_atom')
|
| 276 |
+
# Extract the padding token ID from tokenizer
|
| 277 |
+
image_padding_token_id = self.get_token_value('image_pad')
|
| 278 |
+
|
| 279 |
+
# Create a new list with padding tokens inserted
|
| 280 |
+
padded_placeholder_tokens = []
|
| 281 |
+
for token in image_placeholders:
|
| 282 |
+
padded_placeholder_tokens.append(image_padding_token_id)
|
| 283 |
+
if token == image_atom_token_id:
|
| 284 |
+
padded_placeholder_tokens.extend([image_padding_token_id] * self.image_segment_len)
|
| 285 |
+
return padded_placeholder_tokens
|
| 286 |
+
|
| 287 |
+
def preprocess_image(self, image: PIL.Image.Image, max_partition, covering_threshold, convert_to_rgb, return_tensors):
|
| 288 |
+
def _preprocess(img: PIL.Image.Image, side):
|
| 289 |
+
# first resize and preprocess
|
| 290 |
+
w, h = img.size
|
| 291 |
+
if w == h:
|
| 292 |
+
new_width = new_height = side
|
| 293 |
+
elif w > h:
|
| 294 |
+
new_width = side
|
| 295 |
+
new_height = int(h / w * new_width)
|
| 296 |
+
else:
|
| 297 |
+
new_height = side
|
| 298 |
+
new_width = int(w / h * new_height)
|
| 299 |
+
new_size = dict(height=new_height, width=new_width)
|
| 300 |
+
pixel_values = self.image_processor.preprocess(img, size=new_size, return_tensors=return_tensors)['pixel_values']
|
| 301 |
+
|
| 302 |
+
# then pad to square
|
| 303 |
+
square_values = torch.zeros([1, 3, side, side], dtype=pixel_values.dtype, device=pixel_values.device)
|
| 304 |
+
new_height, new_width = pixel_values.shape[2:]
|
| 305 |
+
if new_height == new_width:
|
| 306 |
+
square_values[:, :, :, :] = pixel_values
|
| 307 |
+
elif new_height > new_width:
|
| 308 |
+
from_index = (side - new_width) // 2
|
| 309 |
+
square_values[:, :, :, from_index:from_index + new_width] = pixel_values
|
| 310 |
+
else:
|
| 311 |
+
from_index = (side - new_height) // 2
|
| 312 |
+
square_values[:, :, from_index:from_index + new_height, :] = pixel_values
|
| 313 |
+
|
| 314 |
+
return square_values
|
| 315 |
+
|
| 316 |
+
def _partition(img, grid) -> list[tuple[int, int, int, int]]:
|
| 317 |
+
w, h = img.size
|
| 318 |
+
row_height = h // grid[0]
|
| 319 |
+
col_width = w // grid[1]
|
| 320 |
+
|
| 321 |
+
partition = []
|
| 322 |
+
for row in range(grid[0]):
|
| 323 |
+
for col in range(grid[1]):
|
| 324 |
+
left = col * col_width
|
| 325 |
+
upper = row * row_height
|
| 326 |
+
right = w if col == grid[1] - 1 else (col + 1) * col_width
|
| 327 |
+
lower = h if row == grid[0] - 1 else (row + 1) * row_height
|
| 328 |
+
partition.append((left, upper, right, lower))
|
| 329 |
+
|
| 330 |
+
return partition
|
| 331 |
+
|
| 332 |
+
def _covering_area(left, upper, right, lower, side):
|
| 333 |
+
w = right - left
|
| 334 |
+
h = lower - upper
|
| 335 |
+
w, h = max(w, h), min(w, h)
|
| 336 |
+
if w > side:
|
| 337 |
+
h = h / w * side
|
| 338 |
+
w = side
|
| 339 |
+
return w * h
|
| 340 |
+
|
| 341 |
+
def _get_best_grid(img, side):
|
| 342 |
+
img_area = img.size[0] * img.size[1]
|
| 343 |
+
|
| 344 |
+
candidate_grids = []
|
| 345 |
+
for i in range(1, max_partition + 1):
|
| 346 |
+
for j in range(1, max_partition + 1):
|
| 347 |
+
if i * j <= max_partition:
|
| 348 |
+
candidate_grids.append((i, j))
|
| 349 |
+
|
| 350 |
+
all_grids = []
|
| 351 |
+
good_grids = []
|
| 352 |
+
for grid in candidate_grids:
|
| 353 |
+
partition = _partition(img, grid)
|
| 354 |
+
covering_ratio = sum([_covering_area(*p, side) for p in partition]) / img_area
|
| 355 |
+
assert covering_ratio <= 1.0
|
| 356 |
+
all_grids.append((grid, covering_ratio))
|
| 357 |
+
if covering_ratio > covering_threshold:
|
| 358 |
+
good_grids.append((grid, covering_ratio))
|
| 359 |
+
|
| 360 |
+
if len(good_grids) > 0:
|
| 361 |
+
# pick the good partition with minimum #sub_images and break the tie using covering_ratio
|
| 362 |
+
return sorted(good_grids, key=lambda x: (x[0][0] * x[0][1], -x[1]))[0][0]
|
| 363 |
+
else:
|
| 364 |
+
# pick the partition with maximum covering_ratio and break the tie using #sub_images
|
| 365 |
+
return sorted(all_grids, key=lambda x: (-x[1], x[0][0] * x[0][1]))[0][0]
|
| 366 |
+
|
| 367 |
+
if convert_to_rgb:
|
| 368 |
+
image = convert_image_mode(image, 'RGB')
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
sides = self.get_image_size()
|
| 372 |
+
if sides[0] != sides[1]:
|
| 373 |
+
raise ValueError('get_image_size() returns non-square size')
|
| 374 |
+
side = sides[0]
|
| 375 |
+
grid = _get_best_grid(image, side)
|
| 376 |
+
partition = _partition(image, grid)
|
| 377 |
+
crops = [image.crop(p) for p in partition]
|
| 378 |
+
if len(crops) > 1:
|
| 379 |
+
crops.insert(0, image)
|
| 380 |
+
pixel_values = torch.cat([_preprocess(crop, side) for crop in crops], dim=0)
|
| 381 |
+
image_placeholders = self.construct_image_placeholders(grid)
|
| 382 |
+
return pixel_values, image_placeholders, grid
|
| 383 |
+
|
| 384 |
+
def batch_decode(self, *args, **kwargs):
|
| 385 |
+
"""
|
| 386 |
+
This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
|
| 387 |
+
refer to the docstring of this method for more information.
|
| 388 |
+
"""
|
| 389 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
| 390 |
+
|
| 391 |
+
def decode(self, *args, **kwargs):
|
| 392 |
+
"""
|
| 393 |
+
This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
|
| 394 |
+
the docstring of this method for more information.
|
| 395 |
+
"""
|
| 396 |
+
return self.tokenizer.decode(*args, **kwargs)
|
| 397 |
+
|
| 398 |
+
def post_process_image_text_to_text(self, generated_outputs):
|
| 399 |
+
"""
|
| 400 |
+
Post-process the output of the model to decode the text.
|
| 401 |
+
Args:
|
| 402 |
+
generated_outputs (`torch.Tensor` or `np.ndarray`):
|
| 403 |
+
The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
|
| 404 |
+
or `(sequence_length,)`.
|
| 405 |
+
Returns:
|
| 406 |
+
`list[str]`: The decoded text.
|
| 407 |
+
"""
|
| 408 |
+
return self.tokenizer.batch_decode(
|
| 409 |
+
generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
| 410 |
+
)
|
| 411 |
+
|
| 412 |
+
@property
|
| 413 |
+
def model_input_names(self):
|
| 414 |
+
tokenizer_input_names = self.tokenizer.model_input_names
|
| 415 |
+
image_processor_input_names = self.image_processor.model_input_names
|
| 416 |
+
names_from_processor = list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
| 417 |
+
return names_from_processor + ["second_per_grid_ts"]
|
| 418 |
+
|
| 419 |
+
|
| 420 |
+
AutoProcessor.register("OvisProcessor", OvisProcessor)
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/tokenizers/__init__.py
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
from .mistral import (MistralTokenizer, maybe_serialize_tool_calls,
|
| 5 |
+
truncate_tool_call_ids, validate_request_params)
|
| 6 |
+
|
| 7 |
+
__all__ = [
|
| 8 |
+
"MistralTokenizer", "maybe_serialize_tool_calls", "truncate_tool_call_ids",
|
| 9 |
+
"validate_request_params"
|
| 10 |
+
]
|
tool_server/.venv/lib/python3.12/site-packages/vllm/transformers_utils/tokenizers/mistral.py
ADDED
|
@@ -0,0 +1,516 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
import os
|
| 5 |
+
from pathlib import Path
|
| 6 |
+
from typing import TYPE_CHECKING, Any, Optional, Union, cast
|
| 7 |
+
|
| 8 |
+
import huggingface_hub
|
| 9 |
+
import regex as re
|
| 10 |
+
from huggingface_hub import HfApi, hf_hub_download
|
| 11 |
+
from transformers.tokenization_utils_base import BatchEncoding
|
| 12 |
+
|
| 13 |
+
from vllm.logger import init_logger
|
| 14 |
+
from vllm.transformers_utils.tokenizer_base import TokenizerBase
|
| 15 |
+
from vllm.utils import is_list_of
|
| 16 |
+
|
| 17 |
+
if TYPE_CHECKING:
|
| 18 |
+
# make sure `mistral_common` is lazy imported,
|
| 19 |
+
# so that users who only use non-mistral models
|
| 20 |
+
# will not be bothered by the dependency.
|
| 21 |
+
from mistral_common.protocol.instruct.request import ChatCompletionRequest
|
| 22 |
+
from mistral_common.tokens.tokenizers.mistral import (
|
| 23 |
+
MistralTokenizer as PublicMistralTokenizer)
|
| 24 |
+
|
| 25 |
+
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
|
| 26 |
+
|
| 27 |
+
logger = init_logger(__name__)
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
def maybe_serialize_tool_calls(request: "ChatCompletionRequest"):
|
| 31 |
+
# SEE: https://github.com/vllm-project/vllm/pull/9951
|
| 32 |
+
# Credits go to: @gcalmettes
|
| 33 |
+
# NOTE: There is currently a bug in pydantic where attributes
|
| 34 |
+
# declared as iterables are replaced in in the instances by
|
| 35 |
+
# pydantic-core ValidatorIterator instance. In particular, this
|
| 36 |
+
# affects tool_calls defined in ChatCompletionAssistantMessageParam
|
| 37 |
+
# model:
|
| 38 |
+
# see:
|
| 39 |
+
# - https://github.com/pydantic/pydantic/issues/9467
|
| 40 |
+
# As a result, tool_calls from assistant messages are never
|
| 41 |
+
# deserialized in the request object if the tool_calls iterator is
|
| 42 |
+
# not consumed. This affect messages passed to the MistralTokenizer
|
| 43 |
+
# since no chat template is applied and therefore the tools_calls
|
| 44 |
+
# iterator is not directly consumed.
|
| 45 |
+
# Issue is tracked on Pydantic side, with resolution planned for
|
| 46 |
+
# v2.11 release. In the meantime, the official workaround is to
|
| 47 |
+
# consume the iterator so the tool_calls are correctly deserialized
|
| 48 |
+
# in the OpenAI ChatCompletionAssistantMessageParam object
|
| 49 |
+
# https://github.com/pydantic/pydantic/issues/9467#issuecomment-2442097291 # noqa: E501
|
| 50 |
+
# Official Pydantic Issues:
|
| 51 |
+
# - https://github.com/pydantic/pydantic/issues/9541
|
| 52 |
+
# TODO: remove when pydantic v2.11 is released
|
| 53 |
+
for i, message in enumerate(request.messages):
|
| 54 |
+
if message.get("role") == 'assistant':
|
| 55 |
+
tool_calls_validator = message.get("tool_calls", ().__iter__())
|
| 56 |
+
validated_tool_calls = []
|
| 57 |
+
while True:
|
| 58 |
+
try:
|
| 59 |
+
tool_call = next(tool_calls_validator) # type: ignore
|
| 60 |
+
validated_tool_calls.append(tool_call)
|
| 61 |
+
except StopIteration:
|
| 62 |
+
break
|
| 63 |
+
|
| 64 |
+
request.messages[i]["tool_calls"] = validated_tool_calls
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def truncate_tool_call_ids(request: "ChatCompletionRequest"):
|
| 68 |
+
"""Truncates tool call IDs for Mistral's ID requirements."""
|
| 69 |
+
for i, message in enumerate(request.messages):
|
| 70 |
+
if message.get("role") == 'assistant':
|
| 71 |
+
tool_calls = message.get("tool_calls", [])
|
| 72 |
+
for tool_call in tool_calls:
|
| 73 |
+
if len(tool_call["id"]) > 9:
|
| 74 |
+
logger.warning(
|
| 75 |
+
"Truncating tool call ID: %s to %s",
|
| 76 |
+
tool_call["id"],
|
| 77 |
+
tool_call["id"][-9:],
|
| 78 |
+
)
|
| 79 |
+
tool_call["id"] = tool_call["id"][-9:]
|
| 80 |
+
|
| 81 |
+
request.messages[i]["tool_calls"] = tool_calls
|
| 82 |
+
|
| 83 |
+
elif message.get("role") in {"tool_results", "tool"}:
|
| 84 |
+
if "tool_call_id" in message:
|
| 85 |
+
tool_call_id = message["tool_call_id"]
|
| 86 |
+
|
| 87 |
+
if len(tool_call_id) > 9:
|
| 88 |
+
logger.warning(
|
| 89 |
+
"Truncating tool_call_id: %s to %s",
|
| 90 |
+
tool_call_id,
|
| 91 |
+
tool_call_id[-9:],
|
| 92 |
+
)
|
| 93 |
+
tool_call_id = tool_call_id[-9:]
|
| 94 |
+
request.messages[i]["tool_call_id"] = tool_call_id
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def validate_request_params(request: "ChatCompletionRequest"):
|
| 98 |
+
if (request.skip_special_tokens is not None
|
| 99 |
+
and not request.skip_special_tokens):
|
| 100 |
+
raise ValueError("skip_special_tokens=False is not supported "
|
| 101 |
+
"for Mistral tokenizers.")
|
| 102 |
+
|
| 103 |
+
|
| 104 |
+
def list_local_repo_files(repo_id: str, revision: Optional[str]) -> list[str]:
|
| 105 |
+
repo_cache = os.path.join(
|
| 106 |
+
huggingface_hub.constants.HF_HUB_CACHE,
|
| 107 |
+
huggingface_hub.constants.REPO_ID_SEPARATOR.join(
|
| 108 |
+
["models", *repo_id.split("/")]))
|
| 109 |
+
|
| 110 |
+
if revision is None:
|
| 111 |
+
revision_file = os.path.join(repo_cache, "refs", "main")
|
| 112 |
+
if os.path.isfile(revision_file):
|
| 113 |
+
with open(revision_file) as file:
|
| 114 |
+
revision = file.read()
|
| 115 |
+
|
| 116 |
+
if revision:
|
| 117 |
+
revision_dir = os.path.join(repo_cache, "snapshots", revision)
|
| 118 |
+
if os.path.isdir(revision_dir):
|
| 119 |
+
return os.listdir(revision_dir)
|
| 120 |
+
|
| 121 |
+
return []
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def find_tokenizer_file(files: list[str]):
|
| 125 |
+
file_pattern = re.compile(
|
| 126 |
+
r"^tokenizer\.model\.v.*$|^tekken\.json$|^tokenizer\.mm\.model\.v.*$")
|
| 127 |
+
|
| 128 |
+
matched_files = [file for file in files if file_pattern.match(file)]
|
| 129 |
+
if len(matched_files) > 1:
|
| 130 |
+
raise OSError(
|
| 131 |
+
f"Found {len(matched_files)} files matching the "
|
| 132 |
+
f"pattern: `{file_pattern.pattern}`. Make sure only one Mistral "
|
| 133 |
+
f"tokenizer is present in {files}.")
|
| 134 |
+
elif len(matched_files) == 0:
|
| 135 |
+
raise OSError(
|
| 136 |
+
f"Found {len(matched_files)} files matching the "
|
| 137 |
+
f"pattern: `{file_pattern.pattern}`. Make sure that a Mistral "
|
| 138 |
+
f"tokenizer is present in {files}.")
|
| 139 |
+
|
| 140 |
+
return matched_files[0]
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def _aggregate_content(content: list) -> list[dict[str, Any]]:
|
| 144 |
+
aggregated_content: list[dict[str, Any]] = []
|
| 145 |
+
for chunk in content:
|
| 146 |
+
if chunk.get("type"
|
| 147 |
+
) == "text" and aggregated_content and aggregated_content[
|
| 148 |
+
-1].get("type") == "text":
|
| 149 |
+
aggregated_content[-1]["text"] += "\n\n" + chunk.get("text")
|
| 150 |
+
else:
|
| 151 |
+
aggregated_content.append(chunk)
|
| 152 |
+
if len(aggregated_content) == 1 and aggregated_content[0].get(
|
| 153 |
+
"type") == "text":
|
| 154 |
+
content = aggregated_content[0]["text"]
|
| 155 |
+
return content
|
| 156 |
+
|
| 157 |
+
|
| 158 |
+
def make_mistral_chat_completion_request(
|
| 159 |
+
messages: list["ChatCompletionMessageParam"],
|
| 160 |
+
tools: Optional[list[dict[str,
|
| 161 |
+
Any]]] = None) -> "ChatCompletionRequest":
|
| 162 |
+
last_message = cast(dict[str, Any], messages[-1])
|
| 163 |
+
if last_message["role"] == "assistant":
|
| 164 |
+
last_message["prefix"] = True
|
| 165 |
+
|
| 166 |
+
# mistral-common requires AssistantMessage content to be string [1].
|
| 167 |
+
#
|
| 168 |
+
# [1]: https://github.com/mistralai/mistral-common/blob/f4a06998b75ed78bbf5aaf569590b772ea26c9f6/src/mistral_common/protocol/instruct/messages.py#L80
|
| 169 |
+
for message in messages:
|
| 170 |
+
# Remove reasoning_content as unsupported by Mistral
|
| 171 |
+
_ = message.pop("reasoning_content", None) # type: ignore
|
| 172 |
+
|
| 173 |
+
# Convert list text content to string
|
| 174 |
+
if message.get("role") in ("assistant", "tool"):
|
| 175 |
+
content: Any = message.get("content")
|
| 176 |
+
if isinstance(content, list):
|
| 177 |
+
content = _aggregate_content(content)
|
| 178 |
+
message["content"] = content
|
| 179 |
+
|
| 180 |
+
# The Mistral client, in comparison to the OpenAI client, requires the
|
| 181 |
+
# "parameters" dict and the "description" string to be present
|
| 182 |
+
# even if they are empty.
|
| 183 |
+
if tools:
|
| 184 |
+
for function in [
|
| 185 |
+
tool["function"] for tool in tools
|
| 186 |
+
if tool["type"] == "function"
|
| 187 |
+
]:
|
| 188 |
+
if function.get("parameters") is None:
|
| 189 |
+
function["parameters"] = {}
|
| 190 |
+
if function.get("description") is None:
|
| 191 |
+
function["description"] = ""
|
| 192 |
+
|
| 193 |
+
from mistral_common.protocol.instruct.request import ChatCompletionRequest
|
| 194 |
+
return ChatCompletionRequest(messages=messages,
|
| 195 |
+
tools=tools) # type: ignore[type-var]
|
| 196 |
+
|
| 197 |
+
|
| 198 |
+
class MistralTokenizer(TokenizerBase):
|
| 199 |
+
|
| 200 |
+
def __init__(self, tokenizer: "PublicMistralTokenizer") -> None:
|
| 201 |
+
self.mistral = tokenizer
|
| 202 |
+
self.instruct = tokenizer.instruct_tokenizer
|
| 203 |
+
_mistral_version_str = self.instruct.tokenizer.version.value
|
| 204 |
+
self.version: int = int(_mistral_version_str.split("v")[-1])
|
| 205 |
+
|
| 206 |
+
tokenizer_ = tokenizer.instruct_tokenizer.tokenizer
|
| 207 |
+
from mistral_common.tokens.tokenizers.tekken import (
|
| 208 |
+
SpecialTokenPolicy, Tekkenizer)
|
| 209 |
+
self.is_tekken = isinstance(tokenizer_, Tekkenizer)
|
| 210 |
+
from mistral_common.tokens.tokenizers.sentencepiece import (
|
| 211 |
+
SentencePieceTokenizer)
|
| 212 |
+
self.is_spm = isinstance(tokenizer_, SentencePieceTokenizer)
|
| 213 |
+
if self.is_tekken:
|
| 214 |
+
# Make sure special tokens will not raise
|
| 215 |
+
tokenizer_.special_token_policy = SpecialTokenPolicy.IGNORE
|
| 216 |
+
elif self.is_spm:
|
| 217 |
+
pass
|
| 218 |
+
else:
|
| 219 |
+
raise TypeError(f"Unsupported tokenizer: {type(tokenizer_)}")
|
| 220 |
+
|
| 221 |
+
self._vocab = tokenizer_.vocab()
|
| 222 |
+
# Convert to a dict[str, int] to match protocol, but this is a lossy
|
| 223 |
+
# conversion. There may be multiple token ids that decode to the same
|
| 224 |
+
# string due to partial UTF-8 byte sequences being converted to �
|
| 225 |
+
self._vocab_dict = {
|
| 226 |
+
token: idx
|
| 227 |
+
for idx, token in enumerate(self._vocab)
|
| 228 |
+
}
|
| 229 |
+
self.tokenizer = tokenizer_
|
| 230 |
+
self._max_token_id = self.vocab_size - 1
|
| 231 |
+
|
| 232 |
+
@classmethod
|
| 233 |
+
def from_pretrained(cls,
|
| 234 |
+
path_or_repo_id: str,
|
| 235 |
+
*,
|
| 236 |
+
revision: Optional[str] = None) -> "MistralTokenizer":
|
| 237 |
+
if not Path(path_or_repo_id).exists():
|
| 238 |
+
assert len(path_or_repo_id.split("/")) == 2, (
|
| 239 |
+
"You have either provided a non-existent path: "
|
| 240 |
+
"{path_or_repo_id} or an invalid HF Hub repo id.")
|
| 241 |
+
tokenizer_file = cls._download_mistral_tokenizer_from_hf(
|
| 242 |
+
path_or_repo_id, revision)
|
| 243 |
+
elif Path(path_or_repo_id).is_dir():
|
| 244 |
+
tokenizer_file_name = find_tokenizer_file(
|
| 245 |
+
os.listdir(path_or_repo_id))
|
| 246 |
+
tokenizer_file = str(Path(path_or_repo_id) / tokenizer_file_name)
|
| 247 |
+
else:
|
| 248 |
+
assert Path(
|
| 249 |
+
path_or_repo_id).is_file(), f"Invalid path: {path_or_repo_id}"
|
| 250 |
+
tokenizer_file = str(Path(path_or_repo_id))
|
| 251 |
+
|
| 252 |
+
from mistral_common.tokens.tokenizers.mistral import (
|
| 253 |
+
MistralTokenizer as PublicMistralTokenizer)
|
| 254 |
+
mistral_tokenizer = PublicMistralTokenizer.from_file(tokenizer_file)
|
| 255 |
+
return cls(mistral_tokenizer)
|
| 256 |
+
|
| 257 |
+
@staticmethod
|
| 258 |
+
def _download_mistral_tokenizer_from_hf(tokenizer_name: str,
|
| 259 |
+
revision: Optional[str]) -> str:
|
| 260 |
+
try:
|
| 261 |
+
hf_api = HfApi()
|
| 262 |
+
files = hf_api.list_repo_files(repo_id=tokenizer_name,
|
| 263 |
+
revision=revision)
|
| 264 |
+
except ConnectionError as exc:
|
| 265 |
+
files = list_local_repo_files(repo_id=tokenizer_name,
|
| 266 |
+
revision=revision)
|
| 267 |
+
|
| 268 |
+
if len(files) == 0:
|
| 269 |
+
raise exc
|
| 270 |
+
|
| 271 |
+
filename = find_tokenizer_file(files)
|
| 272 |
+
|
| 273 |
+
tokenizer_file = hf_hub_download(tokenizer_name,
|
| 274 |
+
filename=filename,
|
| 275 |
+
revision=revision)
|
| 276 |
+
return tokenizer_file
|
| 277 |
+
|
| 278 |
+
# the following attributes are set to fit vLLM's design and are used
|
| 279 |
+
# by the guided structured output backends.
|
| 280 |
+
@property
|
| 281 |
+
def all_special_tokens_extended(self) -> list[str]:
|
| 282 |
+
from mistral_common.tokens.tokenizers.base import SpecialTokens
|
| 283 |
+
|
| 284 |
+
# tekken defines its own extended special tokens list
|
| 285 |
+
if hasattr(self.tokenizer, "SPECIAL_TOKENS"):
|
| 286 |
+
special_tokens = self.tokenizer.SPECIAL_TOKENS
|
| 287 |
+
else:
|
| 288 |
+
special_tokens = list(SpecialTokens)
|
| 289 |
+
return [
|
| 290 |
+
s.value if isinstance(s, SpecialTokens) else s
|
| 291 |
+
for s in special_tokens
|
| 292 |
+
]
|
| 293 |
+
|
| 294 |
+
@property
|
| 295 |
+
def all_special_tokens(self) -> list[str]:
|
| 296 |
+
return self.all_special_tokens_extended
|
| 297 |
+
|
| 298 |
+
@property
|
| 299 |
+
def all_special_ids(self) -> list[int]:
|
| 300 |
+
return [
|
| 301 |
+
self.all_special_tokens.index(t) for t in self.all_special_tokens
|
| 302 |
+
]
|
| 303 |
+
|
| 304 |
+
@property
|
| 305 |
+
def bos_token_id(self) -> int:
|
| 306 |
+
return self.tokenizer.bos_id
|
| 307 |
+
|
| 308 |
+
@property
|
| 309 |
+
def eos_token_id(self) -> int:
|
| 310 |
+
return self.tokenizer.eos_id
|
| 311 |
+
|
| 312 |
+
@property
|
| 313 |
+
def sep_token(self) -> str:
|
| 314 |
+
raise NotImplementedError()
|
| 315 |
+
|
| 316 |
+
@property
|
| 317 |
+
def pad_token(self) -> str:
|
| 318 |
+
raise NotImplementedError()
|
| 319 |
+
|
| 320 |
+
@property
|
| 321 |
+
def is_fast(self) -> bool:
|
| 322 |
+
return True
|
| 323 |
+
|
| 324 |
+
@property
|
| 325 |
+
def vocab_size(self) -> int:
|
| 326 |
+
return len(self._vocab)
|
| 327 |
+
|
| 328 |
+
@property
|
| 329 |
+
def max_token_id(self) -> int:
|
| 330 |
+
return self._max_token_id
|
| 331 |
+
|
| 332 |
+
def __len__(self) -> int:
|
| 333 |
+
return self.vocab_size
|
| 334 |
+
|
| 335 |
+
def __call__(
|
| 336 |
+
self,
|
| 337 |
+
text: Union[str, list[str], list[int]],
|
| 338 |
+
text_pair: Optional[str] = None,
|
| 339 |
+
add_special_tokens: bool = False,
|
| 340 |
+
truncation: bool = False,
|
| 341 |
+
max_length: Optional[int] = None,
|
| 342 |
+
):
|
| 343 |
+
input_ids: Union[list[int], list[list[int]]]
|
| 344 |
+
# For list[str], original prompt text
|
| 345 |
+
if is_list_of(text, str):
|
| 346 |
+
input_ids_: list[list[int]] = []
|
| 347 |
+
for p in text:
|
| 348 |
+
each_input_ids = self.encode_one(p, truncation, max_length)
|
| 349 |
+
input_ids_.append(each_input_ids)
|
| 350 |
+
input_ids = input_ids_
|
| 351 |
+
# For list[int], apply chat template output, already tokens.
|
| 352 |
+
elif is_list_of(text, int):
|
| 353 |
+
input_ids = text
|
| 354 |
+
# For str, single prompt text
|
| 355 |
+
else:
|
| 356 |
+
input_ids = self.encode_one(text, truncation, max_length)
|
| 357 |
+
return BatchEncoding({"input_ids": input_ids})
|
| 358 |
+
|
| 359 |
+
def get_vocab(self) -> dict[str, int]:
|
| 360 |
+
# NB: the dictionary form of the vocabulary collapses token ids that map
|
| 361 |
+
# to the same string but have different bytes
|
| 362 |
+
return self._vocab_dict
|
| 363 |
+
|
| 364 |
+
def get_added_vocab(self) -> dict[str, int]:
|
| 365 |
+
# Mistral tokenizers have no added vocabulary
|
| 366 |
+
return {}
|
| 367 |
+
|
| 368 |
+
def encode_one(
|
| 369 |
+
self,
|
| 370 |
+
text: str,
|
| 371 |
+
truncation: bool = False,
|
| 372 |
+
max_length: Optional[int] = None,
|
| 373 |
+
) -> list[int]:
|
| 374 |
+
# Mistral Tokenizers should not add special tokens
|
| 375 |
+
input_ids = self.encode(text)
|
| 376 |
+
|
| 377 |
+
if truncation:
|
| 378 |
+
input_ids = input_ids[:max_length]
|
| 379 |
+
return input_ids
|
| 380 |
+
|
| 381 |
+
def encode(self,
|
| 382 |
+
text: str,
|
| 383 |
+
truncation: Optional[bool] = None,
|
| 384 |
+
max_length: Optional[int] = None,
|
| 385 |
+
add_special_tokens: Optional[bool] = None) -> list[int]:
|
| 386 |
+
# `encode` should only be used for prompt completion
|
| 387 |
+
# it should never be used for chat_completion.
|
| 388 |
+
# For chat completion use `apply_chat_template`
|
| 389 |
+
if add_special_tokens is not None:
|
| 390 |
+
return self.tokenizer.encode(text,
|
| 391 |
+
bos=add_special_tokens,
|
| 392 |
+
eos=add_special_tokens)
|
| 393 |
+
else:
|
| 394 |
+
return self.tokenizer.encode(text, bos=True, eos=False)
|
| 395 |
+
|
| 396 |
+
def apply_chat_template(self,
|
| 397 |
+
messages: list["ChatCompletionMessageParam"],
|
| 398 |
+
tools: Optional[list[dict[str, Any]]] = None,
|
| 399 |
+
**kwargs) -> list[int]:
|
| 400 |
+
|
| 401 |
+
request = make_mistral_chat_completion_request(messages, tools)
|
| 402 |
+
encoded = self.mistral.encode_chat_completion(request)
|
| 403 |
+
|
| 404 |
+
# encode-decode to get clean prompt
|
| 405 |
+
return encoded.tokens
|
| 406 |
+
|
| 407 |
+
def convert_tokens_to_string(self, tokens: list[str]) -> str:
|
| 408 |
+
from mistral_common.tokens.tokenizers.base import SpecialTokens
|
| 409 |
+
if self.is_tekken:
|
| 410 |
+
tokens = [
|
| 411 |
+
t for t in tokens
|
| 412 |
+
if (t is SpecialTokens.tool_calls
|
| 413 |
+
or t not in self.tokenizer._all_special_tokens)
|
| 414 |
+
]
|
| 415 |
+
|
| 416 |
+
if any(isinstance(t, bytes) for t in tokens):
|
| 417 |
+
# we need to encode and decode all tokens again
|
| 418 |
+
shift = self.tokenizer.num_special_tokens
|
| 419 |
+
|
| 420 |
+
def _token_to_id(t: str):
|
| 421 |
+
t_bytes = t.encode("utf-8") \
|
| 422 |
+
if not isinstance(t, bytes) else t
|
| 423 |
+
try:
|
| 424 |
+
return shift + \
|
| 425 |
+
self.tokenizer._tekken_token2id_nospecial[t_bytes]
|
| 426 |
+
except KeyError:
|
| 427 |
+
logger.warning(
|
| 428 |
+
"Failed to convert token %s to id,"
|
| 429 |
+
" replacing with <unk>", t_bytes)
|
| 430 |
+
return self.tokenizer.unk_id
|
| 431 |
+
|
| 432 |
+
ids = [_token_to_id(t) for t in tokens]
|
| 433 |
+
decoded = self.tokenizer.decode(ids)
|
| 434 |
+
else:
|
| 435 |
+
decoded = "".join(tokens)
|
| 436 |
+
else:
|
| 437 |
+
# make sure certain special tokens like Tool calls are
|
| 438 |
+
# not decoded
|
| 439 |
+
special_tokens = {SpecialTokens.tool_calls}
|
| 440 |
+
regular_tokens: list[str] = []
|
| 441 |
+
decoded_list = []
|
| 442 |
+
|
| 443 |
+
for token in tokens:
|
| 444 |
+
if token in special_tokens:
|
| 445 |
+
if regular_tokens:
|
| 446 |
+
decoded_list.append(
|
| 447 |
+
self.tokenizer.decode(regular_tokens))
|
| 448 |
+
regular_tokens = []
|
| 449 |
+
decoded_list.append(token)
|
| 450 |
+
else:
|
| 451 |
+
regular_tokens.append(token)
|
| 452 |
+
|
| 453 |
+
if regular_tokens:
|
| 454 |
+
decoded_list.append(
|
| 455 |
+
self.tokenizer.decode(regular_tokens)) # type: ignore
|
| 456 |
+
|
| 457 |
+
decoded = ''.join(decoded_list)
|
| 458 |
+
|
| 459 |
+
return decoded
|
| 460 |
+
|
| 461 |
+
# WARN: Outlines logits processors can overwrite this method.
|
| 462 |
+
# See: guided_decoding/outlines_logits_processors.py::_adapt_tokenizer
|
| 463 |
+
# for more.
|
| 464 |
+
def decode(self,
|
| 465 |
+
ids: Union[list[int], int],
|
| 466 |
+
skip_special_tokens: bool = True) -> str:
|
| 467 |
+
assert (
|
| 468 |
+
skip_special_tokens
|
| 469 |
+
), "skip_special_tokens=False is not supported for Mistral tokenizers."
|
| 470 |
+
|
| 471 |
+
if isinstance(ids, int):
|
| 472 |
+
ids = [ids]
|
| 473 |
+
return self.tokenizer.decode(ids)
|
| 474 |
+
|
| 475 |
+
def convert_ids_to_tokens(
|
| 476 |
+
self,
|
| 477 |
+
ids: list[int],
|
| 478 |
+
skip_special_tokens: bool = True,
|
| 479 |
+
) -> list[str]:
|
| 480 |
+
from mistral_common.tokens.tokenizers.base import SpecialTokens
|
| 481 |
+
from mistral_common.tokens.tokenizers.instruct import (
|
| 482 |
+
InstructTokenizerV13)
|
| 483 |
+
|
| 484 |
+
# TODO(Patrick) - potentially allow special tokens to not be skipped
|
| 485 |
+
assert (
|
| 486 |
+
skip_special_tokens
|
| 487 |
+
), "skip_special_tokens=False is not supported for Mistral tokenizers."
|
| 488 |
+
|
| 489 |
+
assert self.is_tekken or self.is_spm, type(self.tokenizer)
|
| 490 |
+
|
| 491 |
+
if self.is_tekken:
|
| 492 |
+
# skip special tokens except tool call and think tokens
|
| 493 |
+
non_skip_special_tokens = {
|
| 494 |
+
self.tokenizer.get_control_token(SpecialTokens.tool_calls)
|
| 495 |
+
}
|
| 496 |
+
if isinstance(self.instruct, InstructTokenizerV13):
|
| 497 |
+
if self.instruct.BEGIN_THINK:
|
| 498 |
+
non_skip_special_tokens.add(self.instruct.BEGIN_THINK)
|
| 499 |
+
if self.instruct.END_THINK:
|
| 500 |
+
non_skip_special_tokens.add(self.instruct.END_THINK)
|
| 501 |
+
ids = [
|
| 502 |
+
i for i in ids if i > self.tokenizer.num_special_tokens
|
| 503 |
+
or i in non_skip_special_tokens
|
| 504 |
+
]
|
| 505 |
+
|
| 506 |
+
tokens = [self.tokenizer.id_to_piece(id) for id in ids]
|
| 507 |
+
|
| 508 |
+
if any("�" in t for t in tokens) and self.is_tekken:
|
| 509 |
+
# if a decoded token contains the replacement character, then the
|
| 510 |
+
# token has an incomplete UTF-8 character so we must use bytes
|
| 511 |
+
# See: https://github.com/vllm-project/vllm/pull/8640
|
| 512 |
+
# https://github.com/vllm-project/vllm/pull/9625
|
| 513 |
+
# if underlying tokenizeir is sentencepiece, we just add "�"
|
| 514 |
+
tokens = [self.tokenizer.id_to_byte_piece(id) for id in ids]
|
| 515 |
+
|
| 516 |
+
return tokens
|