Update RTX 5090 Blackwell runner with 2-step setup instructions

runners/RTX_5090_BLACKWELL_PRETRAIN.txt

@@ -1,21 +1,15 @@
 # ==============================================================================
 # 🚀 ViuAI Sarus-500M — NVIDIA RTX 5090 Blackwell Turbo Pretraining (32GB VRAM)
 # ==============================================================================
-# Instructions:
-# 1. Start your RTX 5090 instance on Race Engineering (Delhi Datacenter).
-# 2. Open JupyterLab, paste this entire cell into a fresh notebook, and run (Shift + Enter).
-# 3. Performance Targets:
-#    • Expected Speed: ~36,000 - 45,000 tokens/sec.
-#    • 1 Step Time: ~12 - 14 seconds (524,288 tokens/step).
-#    • 1,000 Steps (~524M tokens): ~3.6 - 4.0 Hours.
-#    • Auto-resumes from Step 40,250 checkpoint on Hugging Face (ViuAI/ViuAI-500M).
-#    • Streams 200 shards (10B tokens) from ViuAI/viuai-500m-data.
-#    • Auto-uploads ckpt_latest.pt to Hugging Face in the background every 250 steps.
-# ==============================================================================
-
-# Ensure PyTorch with CUDA 12.8 (sm_120 Blackwell support) is installed
-!pip install -q --upgrade --index-url https://download.pytorch.org/whl/cu128 torch
-!pip install -q transformers datasets tokenizers huggingface_hub numpy hf_transfer
+# Instructions for JupyterLab on RTX 5090 (Blackwell sm_120):
+# 
+# 📌 CELL 1: Run once to install CUDA 12.8 PyTorch with native sm_120 Blackwell support.
+# ------------------------------------------------------------------------------
+# !pip install --force-reinstall --no-deps -q torch --index-url https://download.pytorch.org/whl/cu128
+# !pip install -q transformers datasets tokenizers huggingface_hub numpy hf_transfer
+# 
+# 📌 CELL 2: Run to launch continuation pretraining to 60,000 steps!
+# ------------------------------------------------------------------------------
 
 import os, sys, shutil
 import torch