Update README.md

README.md

@@ -12,6 +12,9 @@ license: apache-2.0
 pipeline_tag: text-generation
 ---
 
+# Note: If you have a multi-GPU SM120 Blackwell system (RTX 50/Pro), try my vLLM fork to resolve P2P / TP=2 issues (Pending PR into upstream). 
+https://github.com/Gadflyii/vllm/tree/main
+
 # Qwen3-Coder-Next-NVFP4
 
 NVFP4 quantized version of [Qwen/Qwen3-Coder-Next](https://huggingface.co/Qwen/Qwen3-Coder-Next) (80B-A3B).
@@ -32,7 +35,7 @@ NVFP4 quantized version of [Qwen/Qwen3-Coder-Next](https://huggingface.co/Qwen/Q
 
 ## Quantization Details
 
-Quantized using [llmcompressor](https://github.com/vllm-project/llm-compressor) 0.9.0.1 following RedHatAI's approach.
+Quantized using [llmcompressor](https://github.com/vllm-project/llm-compressor) 0.9.0.1.
 
 ```python
 NUM_CALIBRATION_SAMPLES = 20
@@ -60,43 +63,20 @@ ignore = [
 
 ### Context Length Testing
 
-Successfully tested up to **128K tokens** with FP8 KV cache (~95GB VRAM per GPU).
+Successfully tested up to **128K tokens** with FP8 KV cache (Not enough VRAM to test any higher context).
 
 ## Usage with vLLM
 
-Requires vLLM with NVFP4 support (0.16.0+).
+Requires vLLM with NVFP4 support (0.16.0+), Transformers 5.0.0+
 
 ```bash
-# Standard serving (TP=2)
-vllm serve GadflyII/Qwen3-Coder-Next-NVFP4 \
-    --tensor-parallel-size 2 \
-    --max-model-len 49152
-
-# Extended context (128K with FP8 KV cache)
+#vllm Serving
 vllm serve GadflyII/Qwen3-Coder-Next-NVFP4 \
     --tensor-parallel-size 2 \
     --max-model-len 131072 \
     --kv-cache-dtype fp8
-
-# Expert Parallelism (alternative to TP)
-vllm serve GadflyII/Qwen3-Coder-Next-NVFP4 \
-    --tensor-parallel-size 1 \
-    --expert-parallel-size 2 \
-    --max-model-len 131072 \
-    --kv-cache-dtype fp8
 ```
 
-### Notes for Blackwell GPUs (SM120)
-
-- First startup may take longer due to CUDA graph compilation
-- SymmMemCommunicator not supported on device capability 12.0 (handled automatically)
-- Some CUTLASS TMA WS grouped gemm tactics may fail during autotuning (non-fatal)
-
-## Hardware Requirements
-
-- **Minimum**: 2x 48GB GPUs (for inference at shorter context)
-- **Recommended**: 2x 96GB GPUs (for 128K context with FP8 KV cache)
-
 ## License
 
 Apache 2.0 (same as base model)