feat: add RTX 5090 & 4090 configs (docs/RTX_CLOUD_RUN.md)

docs/RTX_CLOUD_RUN.md

@@ -0,0 +1,69 @@
+# RTX 5090 & RTX 4090 CLOUD PRETRAINING GUIDE
+> Dedicated production setup for high-speed training of ViuMini-MoE-242M.
+> Native BF16 + FlashAttention-2 + Non-blocking background checkpoint uploads.
+
+---
+
+## 1. Quick Setup (Run on any Cloud Instance / RunPod / Vast.ai / Lambda / Local PC)
+
+Run these 3 lines in your terminal:
+```bash
+# 1. Clone repository from Hugging Face
+git clone https://huggingface.co/ViuAI/ViuMini-MoE-242M
+cd ViuMini-MoE-242M
+
+# 2. Install dependencies
+pip install -q torch pyarrow datasets tokenizers pyyaml huggingface_hub
+
+# 3. Export your Hugging Face token (for auto-checkpoint uploads)
+export HF_TOKEN="your_huggingface_token_here"
+```
+
+---
+
+## 2. Launch on RTX 5090 (32GB GDDR7) — 1-Click Launch
+
+* **Speed**: **~85,000 – 95,000 tokens/sec**
+* **1 Step Time**: **~1.4 seconds** (Global 128 sequences = 131,072 tokens/step)
+* **1,000 Steps**: **~24 minutes**
+* **10,000 Steps (~1.31B tokens)**: **~4.0 hours** (Total cost @ ₹92/h: **₹368**)
+
+```bash
+python model/scripts/train.py \
+  --config model/configs/train_rtx5090.yaml \
+  --model_config model/configs/viu1_moe_config.yaml \
+  --hf_dataset ViuAI/viu-mini-raw-pretrain \
+  --resume auto
+```
+
+---
+
+## 3. Launch on RTX 4090 (24GB GDDR6X) — 1-Click Launch
+
+* **Speed**: **~50,000 – 58,000 tokens/sec**
+* **1 Step Time**: **~2.5 seconds** (Global 128 sequences = 131,072 tokens/step)
+* **1,000 Steps**: **~40 minutes**
+* **10,000 Steps (~1.31B tokens)**: **~6.8 hours** (Total cost @ ₹69/h: **₹462**)
+
+```bash
+python model/scripts/train.py \
+  --config model/configs/train_rtx4090.yaml \
+  --model_config model/configs/viu1_moe_config.yaml \
+  --hf_dataset ViuAI/viu-mini-raw-pretrain \
+  --resume auto
+```
+
+---
+
+## 4. Key Highlights & Safety Features
+
+1. **Auto-Resume (`--resume auto`)**:
+   * If a checkpoint exists locally in `model/checkpoints/`, it resumes immediately.
+   * If running on a fresh cloud instance, it queries Hugging Face Hub (`ViuAI/ViuMini-MoE-242M`), downloads the highest step checkpoint automatically, and resumes seamlessly!
+2. **Global Batch Parity (131,072 tokens/step)**:
+   * Both configs lock `target_global_batch: 128` (128 sequences x 1024 tokens = 131,072 tokens per step).
+   * Micro-batch: 16 (fits easily in 24GB/32GB VRAM).
+   * Gradient accumulation: 8 (cuts gradient accumulation loops in half compared to T4).
+3. **Non-Blocking Background Checkpoint Uploads**:
+   * Forward/backward passes NEVER pause during multi-gigabyte Hugging Face uploads.
+   * Local disk keeps newest 2 full checkpoints (no disk full crash).