[LLMC] UT test_llmc_format.py::test_llmcompressor_static_fp8_attention_config KeyError: 'weight'

[XuehaoSun]

Error log: https://dev.azure.com/lpot-inc/neural-compressor/_build/results?buildId=69556&view=logs&j=12b2e751-7154-5290-1608-528fefa418c8&t=9b52bf1f-13f5-5471-05d0-3b24c18057e6&l=3573

Root cause: vllm-project/compressed-tensors@2aeeccd

=================================== FAILURES ===================================
________________ test_llmcompressor_static_fp8_attention_config ________________

dataloader = <test.helpers.DataLoader object at 0x7644d9faf980>
tmp_path = PosixPath('/tmp/pytest-of-hostuser/pytest-6/test_llmcompressor_static_fp8_1')

    def test_llmcompressor_static_fp8_attention_config(dataloader, tmp_path):
        model_name = get_model_path("stas/tiny-random-llama-2")
        autoround = AutoRound(
            model_name,
            scheme="FP8_STATIC",
            seqlen=8,
            nsamples=2,
            iters=0,
            dataset=dataloader,
            static_attention_dtype="fp8",
        )
        _, quantized_model_path = autoround.quantize_and_save(tmp_path, format="llm_compressor")
    
        with open(os.path.join(quantized_model_path, "config.json")) as f:
            saved_config = json.load(f)
        saved_groups = saved_config["quantization_config"]["config_groups"]
        attention_group = None
        for group in saved_groups.values():
            if "Linear" not in group["targets"]:
                attention_group = group
                break
    
        assert attention_group is not None
        assert attention_group["weights"] is None
        assert attention_group["input_activations"]["num_bits"] == 8
        assert attention_group["input_activations"]["type"] == "float"
        assert attention_group["input_activations"]["strategy"] == "tensor"
        assert attention_group["input_activations"]["dynamic"] is False
        assert attention_group["input_activations"]["symmetric"] is True
        assert saved_config["quantization_config"]["kv_cache_scheme"] is not None
    
>       model = AutoModelForCausalLM.from_pretrained(quantized_model_path, torch_dtype="auto", trust_remote_code=True)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

test_cpu/export/test_llmc_format.py:310: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
../auto_round/utils/common.py:140: in patched
    return underlying_func(klass, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
/home/hostuser/.venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py:402: in from_pretrained
    return model_class.from_pretrained(
/home/hostuser/.venv/lib/python3.12/site-packages/transformers/modeling_utils.py:4416: in from_pretrained
    hf_quantizer.preprocess_model(
/home/hostuser/.venv/lib/python3.12/site-packages/transformers/quantizers/base.py:171: in preprocess_model
    self._process_model_before_weight_loading(model, **kwargs)
/home/hostuser/.venv/lib/python3.12/site-packages/transformers/quantizers/quantizer_compressed_tensors.py:72: in _process_model_before_weight_loading
    self.compressor.compress_model(model=model)
/home/hostuser/.venv/lib/python3.12/site-packages/compressed_tensors/compressors/model_compressors/model_compressor.py:169: in compress_model
    compress_module(module, self.force_compression_format)
/home/hostuser/.venv/lib/python3.12/site-packages/compressed_tensors/compressors/base.py:193: in compress_module
    compressor.compress_module(module)
/home/hostuser/.venv/lib/python3.12/site-packages/compressed_tensors/compressors/base.py:109: in compress_module
    compressed_state_dict = cls.compress(state_dict, scheme)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

cls = <class 'compressed_tensors.compressors.naive_quantized.base.FloatQuantizationCompressor'>
state_dict = {'q_scale': tensor(..., device='meta', size=(1,), dtype=torch.bfloat16), 'k_scale': tensor(..., device='meta', size=(1,), dtype=torch.bfloat16), 'v_scale': tensor(..., device='meta', size=(1,), dtype=torch.bfloat16)}
scheme = QuantizationScheme(targets=['LlamaAttention'], weights=None, input_activations=QuantizationArgs(num_bits=8, type='floa...s_minmax', observer_kwargs={}), output_activations=None, format=<CompressionFormat.float_quantized: 'float-quantized'>)

    @classmethod
    def compress(
        cls, state_dict: TensorStateDict, scheme: QuantizationScheme
    ) -> TensorStateDict:
        """
        Compress a per-module state dict.
    
        Quantizes the weight to the dtype specified by the scheme's
        QuantizationArgs. Handles block quantization padding if needed.
    
        :param state_dict: local-name state dict (weight, weight_scale, …)
        :param scheme: quantization scheme for the weight
        :return: compressed state dict
        """
        state_dict = state_dict.copy()
>       weight = state_dict.pop("weight")
                 ^^^^^^^^^^^^^^^^^^^^^^^^
E       KeyError: 'weight'

/home/hostuser/.venv/lib/python3.12/site-packages/compressed_tensors/compressors/naive_quantized/base.py:63: KeyError