Bump PyTorch to 2.7.0

[AlexanderDokuchaev]

Changes

Update pytorch to 2.7.0
Regenerate ref data for test_generate_text_data_functional, changed output of hf-internal-testing/tiny-random-gpt2

Tests

manual/job/post_training_quantization/661/ - fail FX on save compressed models
nightly/job/TriggerBetta/1029/ -
examples :

• llm_compression_qat_with_lora assert 0.034 == 0.027 ± 2.0e-03
• llm_compression_synthetic - AssertionError: metric word_count: 81 != 83

wc - pass
Test_Install - pass

[ljaljushkin]

It appears that F.linear() is the source of indeterminism.

When model is on the CUDA, the code below

model = AutoModelForCausalLM.from_pretrained(BASE_TEST_MODEL_ID, device_map="cuda")
torch.use_deterministic_algorithms(True)

leads to an error:

../../env/nncf-py/lib/python3.10/site-packages/torch/nn/modules/linear.py:125: in forward
return F.linear(input, self.weight, self.bias)

RuntimeError: Deterministic behavior was enabled with either torch.use_deterministic_algorithms(True) or at::Context::setDeterministicAlgorithms(true), but this operation is not deterministic because it uses CuBLAS and you have CUDA >= 10.2. To enable deterministic behavior in this case, you must set an environment variable before running your PyTorch application: CUBLAS_WORKSPACE_CONFIG=:4096:8 or CUBLAS_WORKSPACE_CONFIG=:16:8. For more information, go to https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility

Then, with export CUBLAS_WORKSPACE_CONFIG = :4096 :8, the results are the same with torch 2.6.0 and 2.7.0

The same thing may happen with qat-lora sample on CPU.
When I switched to GPU, wwb reference were different:

use_deterministic_algorithms + CUBLAS_WORKSPACE_CONFIG aligns it.

Propose updating references, since with use_deterministic_algorithms the test would be slower, but the results were the same.