Fix: 262K context requires both cards unified, not simultaneous

Both GPUs must be dedicated to the LLM for 262K context window.
Image gen and LLM run one at a time (swap takes seconds).
Simultaneous only possible with smaller context (~64K on one card).

https://claude.ai/code/session_01PtYTPherSJaxDEVPgF6Nxu

docs/gpu-setup-research.md

@@ -264,16 +264,24 @@ The RTX 5000 has **384 Tensor Cores with native FP16** — full SDXL/Flux suppor
 | Flux Dev (high-res / batched) | ~18-24GB | Both cards via NVLink (32GB) |
 | ComfyUI / InvokeAI | Works natively | No `--force-fp32` needed |
 
-**Run both workloads simultaneously:**
+**Important: 262K context requires both cards unified.** You can't split one off for image
-```bash
+gen and keep 262K. It's one task at a time:
-# Option A: Dedicated cards (no model swapping)
-CUDA_VISIBLE_DEVICES=0  # Ollama — coding LLM on GPU 0
-CUDA_VISIBLE_DEVICES=1  # ComfyUI/InvokeAI — image gen on GPU 1
 
-# Option B: Both cards unified for whichever task you're doing
+```bash
-# Switch between LLM (32GB, 262K context) and image gen (32GB, high-res batches)
+# CODING SESSION: Both cards unified → 32GB → 262K context
+ollama run qwen3.5:35b-a3b-q4_K_M   # Uses both GPUs via NVLink
+
+# IMAGE GEN SESSION: Stop LLM, run image gen on one card (16GB is plenty)
+ollama stop                           # Frees VRAM
+comfyui --listen 0.0.0.0             # SDXL/Flux fits easily in 16GB
+
+# Swap takes a few seconds, not simultaneous but not painful
 ```
 
+If you want simultaneous coding + image gen, you'd run a smaller model at shorter context
+on one card (e.g., Qwen3.5-35B-A3B at ~64K on 16GB) and image gen on the other. But for
+full 262K context, both cards must be dedicated to the LLM.
+
 ### Hardware Longevity: 3-5 Years Realistic
 
 - **2026-2027**: Sweet spot. MoE + linear attention models are getting smaller active params.