Revert "[DSv4] Refactor compressor & Fix ROCm compatibility" (#43710)

[vllm-agent]

Revert of #43710

This reverts commit adaa5e4 (merge commit of PR #43710).

Original PR: #43710

Reason: PR #43710 introduced an unconditional import cutlass in vllm/models/deepseek_v4/nvidia/ops/__init__.py which causes ModuleNotFoundError: No module named 'cutlass' on environments where cutlass is not installed. This cascades through get_quantization_config() into any test that initializes a model with quantization configs, causing 20 new test failures across the nightly CI build #68351.

Failed tests (20):

• CPU-Quantization Model Tests
• Fusion and Compile Unit Tests (2xB200)
• Kernels (B200)
• Kernels DeepGEMM Test (H100)
• LoRA TP (Distributed)
• Batch Invariance (B200)
• Batch Invariance (H100)
• Model Runner V2 Spec Decode
• Basic Models Tests (Extra Initialization) 1
• Basic Models Tests (Extra Initialization) 2
• Basic Models Tests (Initialization)
• Distributed Model Tests (2 GPUs)
• Language Models Test (PPL)
• Multi-Modal Models (Extended Generation 1)
• Multi-Modal Processor (CPU)
• PyTorch Fullgraph
• Spec Decode Draft Model
• Spec Decode Draft Model Nightly B200
• Spec Decode Eagle
• Weight Loading Multiple GPU

---

Auto-generated by CI failure analyzer. Build #68351.

[tjtanaa]

@khluu @WoosukKwon @zyongye The PR #43710 seems to fix the cutlass import issue for me on ROCm when running deepseek v4. So it partially fixed the issue.