# AWQ + DFlash on Ampere Hardware

## Date: 2026-08-08

## Objective
Deploy Qwen3.6-35B-A3B with AWQ 4bit quantization and DFlash speculative decoding on NVIDIA Ampere-class GPU (24GB VRAM).

## Scripts Created
- `vllm-awq-qwen36-dflash.sh` — Docker-based AWQ + DFlash
- `vllm-awq-qwen36-dflash-ampere.sh` — Systemd service, Ampere-optimized

## Key Findings

### vLLM CLI Changes (Nightly)
- `--speculative-model` removed; use `--speculative-config` JSON instead
- Format: `--speculative-config '{"method": "dflash", "model": "z-lab/Qwen3.6-35B-A3B-DFlash", "num_speculative_tokens": 4}'`
- `num_speculative_tokens` is required when model is provided

### Memory Issues
- **OOM with 262144 context**: AWQ→Marlin weight conversion requires extra GPU memory
- **Solution**: Lower to 32768 context temporarily
- **CPU offload**: `--cpu-offload-gb 40` + `--cpu-offload-params "experts"` helps

### Conflicts
- **Prefix caching + Mamba cache**: Causes `AssertionError` in `unify_kv_cache_spec_page_size`
- **DFlash speculative decoding**: KV cache assertion bug in nightly vLLM
- Workaround: Disable speculative decoding or prefix caching

### Performance Observations (no speculative decoding)
| Streams | Prompt Speed | Completion Speed |
|---------|-------------|------------------|
| 1 | 4.27 tok/s | 21.42 tok/s |
| 2 | 2.44 tok/s | 12.25 tok/s |
| 3 | 1.35 tok/s | 6.76 tok/s |
| 4 | 1.43 tok/s | 7.19 tok/s |

Note: Speeds decrease with more streams due to KV cache and compute contention.

## Systemd Service
- Service: `vllm-awq-qwen36.service`
- ExecStart: `/home/dank/vllm/vllm-awq-qwen36-dflash-ampere.sh`
- Port: 34401
- Status: Disabled (needs speculative decoding fix)

## Next Steps
1. Fix DFlash speculative decoding (upstream vLLM bug or workaround)
2. Test with lower context length for full functionality
3. Benchmark with and without speculative decoding
4. Investigate `max_num_scheduled_tokens` warning (set to 1530 due to speculative settings)