new installation instuction for stable/correct performance

README.md

@@ -17,17 +17,18 @@ base_model_relation: quantized
 Base model: [DeepSeek-V3.1](https://huggingface.co/deepseek-ai/DeepSeek-V3.1)
 
 ### 【Dependencies / Installation】
-As of **2025-08-23**, create a fresh Python environment and run:
+As of **2025-08-27**, create a fresh Python environment and run:
 
 ```bash
 # ❗there are glitches with vllm 0.10.1.1, still looking for resolutions❗
 # ❗downgrade vllm for now ❗
-pip install vllm==0.9.0
-pip install transformers==4.53
+pip install vllm==0.9.2 transformers==4.53.0
 
-# ❗patch up AWQ MoE quant config, otherwise some modules cannot be properly loaded❗
 SITE_PACKAGES=$(pip -V | awk '{print $4}' | sed 's/\/pip$//')
+# ❗patch up AWQ MoE quant config, otherwise some modules cannot be properly loaded❗
 cp awq_marlin.py "$SITE_PACKAGES/vllm/model_executor/layers/quantization/awq_marlin.py"
+# ❗patch up for fp32 e_score_correction_bias, see https://www.github.com/vllm-project/vllm/pull/23640❗
+cp deepseek_v2.py "$SITE_PACKAGES/vllm/model_executor/models/deepseek_v2.py"
 ```
 
 ### 【vLLM Single Node with 8 GPUs — Startup Command】
@@ -51,6 +52,12 @@ vllm serve \
 
 ### 【Logs】
 ```
+2025-08-27
+1. new installation instuction for stable/correct performance
+  (a) use vllm 0.9.2 instead of 0.9.0:
+      there is unidentified issue with 0.9.0 🥹 which causes numerical error
+  (b) patch up deepseek_v2.py for fp32 e_score_correction_bias
+
 2025-08-23
 1. Initial commit
 ```