Add version banner to patch — prints commit, arch, bugs fixed at startup

Ensures we can always verify what's running inside the container
from the docker logs. No functional changes.

patches/deepseek_v4.py

@@ -1,5 +1,38 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# ==============================================================================
+# DeepSeek V4 NVFP4 Patch — Version Banner (printed at import time)
+# ==============================================================================
+import datetime as _dt
+_git_commit = os.popen("git -C /root/nvidia-meeting/deepseek-v4-quant rev-parse --short HEAD 2>/dev/null || echo 'unknown'").read().strip()
+print(f"""
+{'='*70}
+  DeepSeek V4 NVFP4 Patch
+  {'='*70}
+  Commit:   {_git_commit}
+  Loaded:   {_dt.datetime.now().strftime('%Y-%m-%d %H:%M:%S UTC')}
+  Node:     {os.uname().nodename}
+
+  Architecture:
+    wo_a        → FP8 + DeepGEMM block scale (BMM einsum)
+    wq_b/wo_b   → BF16 (UnquantizedLinearMethod)
+    fused_wqa   → BF16 (stacked q_a + kv, dequantized from NVFP4)
+    compressor  → BF16 (reconstructed from separate kv_proj+gate_proj)
+    shared_exp  → FP8 (Fp8LinearMethod, DeepGEMM)
+    MoE experts → NVFP4 (FusedMoE, FLASHINFER_TRTLLM) — NOT converted
+  
+  Bugs fixed:
+    #1 DeepGEMM sf.dim() — block scale format (deepgemm_post_process)
+    #2 fused_skip_regex — q_b/o_a/o_b scales no longer skipped
+    #3 input_scale — removed from weight dequant (activations only)
+    #4 compressor indexer — sub_path for .indexer keys
+    #5 block scale dtype — must be float32, not float8_e4m3fn
+    #6 block scale values — torch.full(fp8_scale) not torch.ones
+{'='*70}
+""")
+# ==============================================================================
+
 import typing
 from collections.abc import Callable, Iterable
 from itertools import islice