Qwen3.8-Flash-Next-AWQ-INT4-FP8PLE

Qwen3.8-Flash-Next-AWQ-INT4-FP8PLE

AWQ-INT4 quant of Qwen/Qwen3.8-Flash-Next for consumer multi-GPU serving on the wtdcode/vllm-backport fork, with the 51.2B-parameter n-gram PLE lookup table converted BF16 → FP8 e4m3. The PLE conversion is what makes CPU-offloaded PLE practical: with VLLM_PLE_CPU_OFFLOAD=1 the pinned-RAM footprint drops from ~95 GiB to ~48 GiB.

• Total size: 127.72 GiB, 38 safetensors shards
• Experts: compressed-tensors AWQ INT4, 4-bit asymmetric, group 32 (routed experts only; attention, norms, MTP, and the PLE table were kept high-precision at quant time)
• PLE table: 128 shard tensors FP8 e4m3 + one global bf16 scale (official Qwen FP8 layout)

Attribution / credits

• Base model: Qwen/Qwen3.8-Flash-Next — Qwen Community License 1.0
• AWQ-INT4 quant: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4 — inherits the base model license
• Serving fork: wtdcode/vllm-backport (docker image lazymio/vllm-backport:latest-sm86) — Apache-2.0. The PLE CPU-offload was implemented by huanghaoyan.hhy (Alibaba) within the fork (commit 93a769a97).
• PLE fp8 layout reference: Qwen/Qwen3.8-Flash-Next-FP8 — Qwen Community License 1.0
• Conversion tooling: serving/convert_fp8_ple.py (Jon-Nielsen)

Licenses

• Model weights (all *.safetensors, config, tokenizer files in this repository): Qwen Community License 1.0 — LICENSE, verbatim from the official Qwen/Qwen3.8-Flash-Next-FP8 repository. Beyond the permissive grant, two conditions apply: (1) using the model in a commercial product or service with more than 100,000,000 monthly active users or US$20,000,000 monthly revenue requires prominently displaying the model name; (2) offering the model (or derivatives) as a commercial "Model as a Service" or "AI Work Assistant" requires a separate license from Qwen — internal use is exempt.
• Code artifacts in serving/: no additional license terms are claimed here. The ple_layer.py overlay is a derivative of vLLM and remains Apache-2.0 (text included as reference: serving/LICENSE). The conversion and verification scripts are original work offered as-is. The upstream licenses above govern the model and everything derived from it.

Conversion details

• 128/128 tensors ...layers.1.ple.ple_embedding.ngram_embedding.shard_{0..127}.weight → float8_e4m3fn
• Single global scale: ...ngram_embedding.weight_scale, bf16, shape (1,), value 0.00019931793212890625, stored in model-00002-of-00038.safetensors
• Raw amax 0.08935546875; the stored scale is the bf16-rounded amax/448, so dequant is exact
• Verified (serving/verify_fp8_ple.py): index ↔ file exact in both directions, zero non-PLE dtype changes, shard_0 SNR 31.5 dB, max relative error 0.195
• Idempotent: atomic per-file writes, resumable, never deletes; reruns skip completed outputs

Serving

Verified on 8× RTX 3090 24 GB (PCIe, no NVLink). serving/docker-compose.yml runs TP=8 + expert parallel (Marlin WNA16 MoE), MTP-3 speculative decoding (4 tokens/step), FULL_AND_PIECEWISE cudagraphs, native 262144 context, --mamba-cache-mode align.

Observed with this exact stack: ~100 tok/s decode; PLE pin ~48 GiB in ~57 s; 582,370-token KV pool at 0.9 gpu-memory-utilization (≈2.2× concurrency at full context).

Quickstart

cd serving
cp .env.example .env    # edit MODEL_PARENT / MODEL / VLLM_API_KEY
docker compose up -d

The overlay bind-mount (serving/overlay/ple_layer.py) is required for this checkpoint: it opts fp8-serialized PLE shards into compressed-tensors checkpoints (VLLM_PLE_FP8_EMBEDDING=1). Without the flag the overlay is inert; without the overlay the flag fails loudly at load time.

KV cache recommendation

Use the default bf16 KV. At 0.85 utilization the rig already holds ~3M tokens of KV across 8 cards, so lower-precision KV buys little headroom here.

Known constraints (fork-specific, verified)

1. Any KV connector combined with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True is rejected by the fork's config validator (blanket rule — the validator cannot enumerate which connectors pin KV memory). Drop the env var when adding a connector.
2. The in-process LMCacheConnectorV1 path has no mamba step-alignment validation. The multi-process sidecar path does: with speculative decoding on, it requires --max-num-batched-tokens equal to the model block size (400) — a fail-closed correctness gate, not a tuning knob.
3. --disable-custom-all-reduce is expected on PCIe-only rigs at TP=8 (vLLM disables custom all-reduce itself above 2 ranks without NVLink); NCCL_ALGO=Ring + NCCL_PROTO=Simple are the fork README's Ampere full-graph prerequisites.

Reproducing the PLE conversion

python serving/convert_fp8_ple.py     # SRC_SNAP / OUT_DIR via env or defaults
python serving/verify_fp8_ple.py      # read-only verification of the output