Update MEMORY.md and memory files with vLLM NVFP4 serving progress

Server running on B200 port 8000 with full NVFP4→vLLM bridge.
All critical bugs fixed: DeepGEMM scale format, compressor shapes, block scale values.

tmp/fix10_quant_method.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+"""Fix the FP8 conversion to use a simple no-op quant method for attention layers."""
+
+filepath = "/root/nvidia-meeting/deepseek-v4-quant/patches/deepseek_v4.py"
+
+with open(filepath, 'r') as f:
+    c = f.read()
+
+# Replace all instances of Fp8LinearMethod/Fp8MMQuantMethod imports and usage
+# with a simpler approach: just set quant_method to None and handle it
+
+# In _convert_nvfp4_module_to_fp8
+old_fp8_convert = '''        # Switch quant method to FP8 linear
+        from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
+        from vllm.model_executor.layers.quantization.utils.quant_utils import (
+            Fp8MMQuantMethod,
+        )
+        mod.quant_method = Fp8LinearMethod(Fp8MMQuantMethod())'''
+
+new_fp8_convert = '''        # Switch quant method to a no-op. The attention forward uses
+        # deepseek_v4_fp8_einsum directly (not the quant method), so the
+        # quant method is irrelevant. We just need process_weights_after_loading
+        # to not crash. Using UnquantizedLinearMethod as a safe no-op.
+        from vllm.model_executor.layers.linear import UnquantizedLinearMethod
+        mod.quant_method = UnquantizedLinearMethod()'''
+
+c = c.replace(old_fp8_convert, new_fp8_convert)
+
+# In the bf16->uint8 handler (o_a_proj case)
+old_oa_fp8 = '''                        # Switch quant method to FP8 linear
+                        from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
+                        from vllm.model_executor.layers.quantization.utils.quant_utils import (
+                            Fp8MMQuantMethod,
+                        )
+                        mod.quant_method = Fp8LinearMethod(Fp8MMQuantMethod())'''
+
+new_oa_fp8 = '''                        # Switch quant method to no-op (attention forward bypasses it)
+                        from vllm.model_executor.layers.linear import UnquantizedLinearMethod
+                        mod.quant_method = UnquantizedLinearMethod()'''
+
+c = c.replace(old_oa_fp8, new_oa_fp8)
+
+with open(filepath, 'w') as f:
+    f.write(c)
+
+print("Replaced Fp8LinearMethod with UnquantizedLinearMethod for attention")