search

huggingface / optimum-intel

🤗 Optimum Intel: Accelerate inference with Intel optimization tools
https://huggingface.co/docs/optimum/main/en/intel/index
Apache License 2.0
364 stars 101 forks source link

issues with IPEXModel.from_pretrained with sentence-transformer models (all-MiniLM-L6-v2, intfloat/e5-mistral-7b-instruct) #810

Closed rbrugaro closed 2 weeks ago

rbrugaro commented 3 weeks ago

I tried 2 different sentence transformers models and during the conversion to torchscript model they fail. I've applied the patch for issue #797 already.

#usage code from https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F
from optimum.intel import IPEXModel

#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Sentences we want sentence embeddings for
sentences = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
#model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2') #works fine
model = IPEXModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2', export=True)

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

# Normalize embeddings
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)

print("Sentence embeddings HF tranformers:")
print(sentence_embeddings)

error: free(): corrupted unsorted chunks

LIBXSMM_VERSION: unconfigured (2147483647) SPR/SP TRY JIT STA COL 0..13 4 4 0 0 14..23 0 0 0 0 24..64 24 24 0 0 Registry and code: 13 MB + 352 KB (gemm=28 gemv=4 meltw=12) Command: python /home/rbrugaro/optimum-intel/notebooks/ipex/test_ST.py Uptime: 0.130897 s Aborted (core dumped)

second model test with usage code from: https://huggingface.co/intfloat/e5-mistral-7b-instruct

import torch
import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel
from optimum.intel import IPEXModel

def last_token_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery: {query}'

# Each query must come with a one-sentence instruction that describes the task
task = 'Given a web search query, retrieve relevant passages that answer the query'
queries = [
    get_detailed_instruct(task, 'how much protein should a female eat'),
    get_detailed_instruct(task, 'summit define')
]
# No need to add instruction for retrieval documents
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]
input_texts = queries + documents

tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-mistral-7b-instruct')
#model = AutoModel.from_pretrained('intfloat/e5-mistral-7b-instruct') #works fine
model = IPEXModel.from_pretrained('intfloat/e5-mistral-7b-instruct')

max_length = 4096
# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')

outputs = model(**batch_dict)
embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])

# normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())

error: $ python /home/rbrugaro/optimum-intel/notebooks/ipex/test_ST2.py /lib/python3.10/site-packages/huggingface_hub/file_download.py:1132: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True. warnings.warn( Detect torchscript is false. Convert to torchscript model! Framework not specified. Using pt to export the model. Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:02<00:00, 1.10s/it] Passing the argument library_name to get_supported_tasks_for_model_type is required, but got library_name=None. Defaulting to transformers. An error will be raised in a future version of Optimum if library_name is not provided. /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/modeling_utils.py:4481: FutureWarning: _is_quantized_training_enabled is going to be deprecated in transformers 4.39.0. Please use model.hf_quantizer.is_trainable instead warnings.warn( /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/modeling_attn_mask_utils.py:276: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! elif sliding_window is None or key_value_length < sliding_window: /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/modeling_attn_mask_utils.py:114: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if (input_shape[-1] > 1 or self.sliding_window is not None) and self.is_causal: /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/modeling_attn_mask_utils.py:162: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if past_key_values_length > 0: /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py:119: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if seq_len > self.max_seq_len_cached: /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py:662: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): Passing the argument library_name to get_supported_tasks_for_model_type is required, but got library_name=None. Defaulting to transformers. An error will be raised in a future version of Optimum if library_name is not provided. /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/amp/autocast_mode.py:267: UserWarning: In CPU autocast, but the target dtype is not supported. Disabling autocast. CPU Autocast only supports dtype of torch.bfloat16, torch.float16 currently. warnings.warn(error_message) Traceback (most recent call last): File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py", line 329, in _call_model out = self.model(*args, kwargs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl return self._call_impl(*args, *kwargs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl return forward_call(args, kwargs) RuntimeError: The following operation failed in the TorchScript interpreter. Traceback of TorchScript (most recent call last): /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py(965): forward /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1522): _slow_forward /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1541): _call_impl /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1532): _wrapped_call_impl /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/jit/_trace.py(1076): trace_module /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/jit/_trace.py(820): trace /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/intel_extension_for_pytorch/jit/_trace.py(69): wrapper /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(118): ipex_jit_trace /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(162): init /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(248): _from_pretrained /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/modeling_base.py(427): from_pretrained /home/rbrugaro/optimum-intel/notebooks/ipex/test_ST2.py(131): RuntimeError: Expected a proper Tensor but got None (or an undefined Tensor in C++) for argument #0 'self'

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "/home/rbrugaro/optimum-intel/notebooks/ipex/test_ST2.py", line 137, in outputs = model(batch_dict) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/modeling_base.py", line 95, in call return self.forward(args, kwargs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py", line 291, in forward outputs = self._call_model(inputs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py", line 331, in _call_model out = self.model(args, kwargs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl return self._call_impl(*args, *kwargs) File "/home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl return forward_call(args, **kwargs) RuntimeError: The following operation failed in the TorchScript interpreter. Traceback of TorchScript (most recent call last): /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py(965): forward /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1522): _slow_forward /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1541): _call_impl /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/nn/modules/module.py(1532): _wrapped_call_impl /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/jit/_trace.py(1076): trace_module /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/torch/jit/_trace.py(820): trace /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/intel_extension_for_pytorch/jit/_trace.py(69): wrapper /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(118): ipex_jit_trace /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(162): init /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/intel/ipex/modeling_base.py(248): _from_pretrained /home/rbrugaro/anaconda3/envs/opti441LANG/lib/python3.10/site-packages/optimum/modeling_base.py(427): from_pretrained /home/rbrugaro/optimum-intel/notebooks/ipex/test_ST2.py(131): RuntimeError: Expected a proper Tensor but got None (or an undefined Tensor in C++) for argument #0 'self'

cc: @jiqing-feng

jiqing-feng commented 2 weeks ago

Hi @rbrugaro . The 1st problem is from tpp, and I have fixed it by #822 .

The 2nd problem is from jit trace, you can see that the jit inputs contains position_ids (see here), so you can add position_ids in your inputs. The correct usage is as follows:

import torch
import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel
from optimum.intel import IPEXModel

from transformers import MistralForCausalLM

def last_token_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery: {query}'

# Each query must come with a one-sentence instruction that describes the task
task = 'Given a web search query, retrieve relevant passages that answer the query'
queries = [
    get_detailed_instruct(task, 'how much protein should a female eat'),
    get_detailed_instruct(task, 'summit define')
]
# No need to add instruction for retrieval documents
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]
input_texts = queries + documents

tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-mistral-7b-instruct')
#model = AutoModel.from_pretrained('intfloat/e5-mistral-7b-instruct') #works fine
model = IPEXModel.from_pretrained('intfloat/e5-mistral-7b-instruct')

max_length = 4096
# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')

position_ids = MistralForCausalLM.prepare_inputs_for_generation(MistralForCausalLM, **batch_dict)["position_ids"]
outputs = model(**batch_dict, position_ids=position_ids)
embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])

# normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())
rbrugaro commented 2 weeks ago

Thanks @jiqing-feng! I verified both issues are fixed