canada-quant/GLM-5.3-Flash-W4A16-MTP

GLM-5.3-Flash — W4A16 (INT4) + BF16 MTP

INT4 weight-only quantization of zai-org/GLM-5.3-Flash with the BF16 MTP draft head kept for speculative decoding. Only the 36,288 routed-expert GEMMs are INT4 (GPTQ, symmetric, group-size 128); attention, router, shared experts, embeddings, the vision tower and the MTP head stay in BF16. Not a new model — all capability comes from the base model.

Full benchmark grids, comparison protocols and research notes: BENCHMARKS.md.

Quick start

# 1. Download (~178 GiB)
huggingface-cli download canada-quant/GLM-5.3-Flash-W4A16-MTP --local-dir /models/glm53-flash-w4a16-mtp

# 2. Serve on 4× H100 / H200 (other hardware: see Serving)
docker run --gpus '"device=0,1,2,3"' --ipc=host --network=host --rm \
  -v /models:/models vllm/vllm-openai:glm53-flash-x86_64-cu130 \
  vllm serve /models/glm53-flash-w4a16-mtp --served-model-name glm53-w4 \
    --tensor-parallel-size 4 --enable-expert-parallel \
    --max-model-len 262144 --max-num-seqs 512 \
    --gpu-memory-utilization 0.92 --no-enable-prefix-caching \
    --speculative-config '{"method":"mtp","num_speculative_tokens":2}' \
    --reasoning-parser glm45 --tool-call-parser glm47 --enable-auto-tool-choice \
    --trust-remote-code --port 8000

# 3. Call it (OpenAI-compatible)
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "glm53-w4",
  "messages": [{"role": "user", "content": "Prove there are infinitely many primes."}]
}'

Two things that bite. If your config.json predates 2026-09-08, re-download it — older copies fail in vLLM with KeyError: 'layers.0.mlp.gate_up_proj.weight' (weights are unchanged). And always pass --max-num-seqs ≤ 512 — the vLLM default of 1024 exceeds this hybrid linear-attention model's 512 Mamba/KDA-state cache blocks.

Hardware and context

Each row is the largest context serving-validated on that configuration.

All configurations use expert parallelism. fp8 KV is not available on Hopper for this NoPE model.

Serving

H100 / H200, TP=4 — the Quick start command is the benchmarked recipe. num_speculative_tokens: 2 is the sweet spot on Hopper (52–55% acceptance; N=5 collapses acceptance to ~30%). Keep prefix caching off — it measured −2…−5% on H200. On 8× H200, two independent TP=4 replicas behind a load balancer beat one TP=8 endpoint by +28–38% aggregate at c128–c512; TP=8 wins single-stream and holds one 7.8M-token pool. On 2× H200 (89.5 GiB weights per GPU) add --tensor-parallel-size 2 --max-num-seqs 128 --max-cudagraph-capture-size 128 for 262K, or --max-model-len 1048576 --max-num-seqs 16 --gpu-memory-utilization 0.95 --max-cudagraph-capture-size 64 --max-num-batched-tokens 4096 for 1M.

RTX PRO 6000, TP=4 — same command with the SM120 image and --max-num-seqs 64 --max-num-batched-tokens 8192 --kv-cache-dtype fp8 --enable-prefix-caching. fp8 KV is required at 262K on 96 GB cards. Keep MTP on at every concurrency here: it adds +70% at c1 and +48% at c32.

2× DGX Spark, TP=2, 1M context — prebuilt image and one-command launcher in canada-quant/vllm-glm53-flash-sm121; the launcher also ships in the drafter repo. Drafter: canada-quant/GLM-5.3-Flash-DFlash2-E (the authors' self-trained DFlash2 drafter, Apache-2.0). Start the worker rank first, wait 25 s, then the head rank.

# on both nodes, rank1 (worker) first, then rank0 (head) 25 s later
MAX_MODEL_LEN=1048576 KV_CACHE_MEM=9663676416 GMU=0.90 EAGER=0 GRAPHS=1 bash launch_dflash2_tp2.sh <rank>

Hard constraints: num_speculative_tokens must be 7 (any other count wedges boot); confirm the boot log shows the mask-embedding load (mask_token_id 154856); keep single prompts ≤ ~310K tokens; stop with docker stop -t 30, never rm -f. Cold boot is 6–10 minutes.

Quality

On RTX PRO 6000, ≈63% of the raw AIME 2025 deficit is a budget wall (empty-answer rate 11.7–14.2% vs 3.3%) and ≈37% is SM120 kernel numerics; a zero-cost commit hook closes it but is not part of the published recipes. The vision tower is BF16 passthrough and was not covered by the text-only calibration; image smoke tests pass, but no formal vision evaluation has been run. Details in BENCHMARKS.md.

Throughput

Output tok/s, thinking ON, same hardware, flags and prompts within each row.

Single-stream decode is insensitive to KV length up to ≥486K (RTX PRO 6000); at batch, long-KV decode plateaus at ~2–4 tok/s per stream and long prefills serialize at a ~6–8.5K tok/s aggregate ceiling. All grids, protocols and the H200 extended table: BENCHMARKS.md.

Known issues

• —`config.json` (2026-09-08): vLLM matches quantization_config.ignore against its own fused module names, so the ignore list now carries both the HF and vLLM spellings plus re:.*\.layers\.45\..* for the MTP head. Older 765-entry copies fail at load. Weights unchanged.
• —DFlash2 admission wedge (SM90 research stack only, MTP recipes unaffected): with the DFlash2 drafter at block size 2304, prompts above ~15.5K tokens are never admitted. A fix was validated to 256K prompts; block size 1536 avoids it. Filed as vllm-project/vllm#55800.
• —Marlin no-split-K path on SM121: deterministic illegal memory access at M=256 when forcing split_k=1; the stock heuristic used in serving is clean. Filed as vllm-project/vllm#56064.

Quantization details

Build gates, all passing: exactly 36,288 packed tensors and nothing quantized outside routed experts; vision key set 348/348 identical to source; MTP layer present; zero dtype drift vs source; no collapsed expert scales. Loads with transformers ≥ 5.16; text generation and image captioning smoke tests pass.

License

MIT, inherited from the base model. Follow the base model's usage terms.

Built, benchmarked and documented with the [Digby.ai](https://digby.ai) coding harness, developed by [CQL.ca](https://cql.ca).