HaberstrohSystems/Qwen3.8-27B-int4-g128-sym-autoround

Qwen3.8-27B — INT4 (AutoRound), tuned for Intel Arc Pro B70

A 27-billion-parameter, natively multimodal model with tool calling, served from a single Intel Arc Pro B70.

INT4 quantization of Qwen/Qwen3.8-27B via AutoRound (W4A16, group size 128, symmetric), packed as auto_round:auto_gptq. Weights land at 17.69 GiB (19.00 GB) across 7 shards plus an extra-tensor file, leaving room on a 32 GB card for a substantial KV cache.

The configuration is the interesting part: Battlemage rewards a different setup than CUDA does, and the settings below are the ones that came out ahead. Every figure here was taken on a single B70, with its measurement conditions stated alongside it.

Read this before you start it

Reasoning is on by default, at the most expensive setting. The chat template resolves reasoning_effort to xhigh unless told otherwise:

{%- if enable_thinking is undefined or enable_thinking is true %}
    {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}

Supported levels are xhigh (default), medium and low. At xhigh, an 8,000-token prompt took 77–100 s before the first word of the answer appeared in our runs — the model writes a long, invisible reasoning pass first. If you start the server with the command below and send a plain request, that is what you get, and it is easy to mistake for a broken setup.

Per request:

{"chat_template_kwargs": {"enable_thinking": false}}
{"chat_template_kwargs": {"reasoning_effort": "low"}}

`preserve_thinking` also defaults to true, so reasoning blocks from earlier turns stay in the prompt and context grows faster than a plain turn count suggests. Set it to false if you want conversations to stay compact.

Sampling: the shipped generation_config.json carries the upstream recommendation for reasoning mode — temperature 1.0, top_p 0.95, top_k 20. For instruct-style use without reasoning, upstream suggests temperature 0.7, top_p 0.80, presence_penalty 1.5.

What you get

What exactly was quantized

block_name_to_quantize is model.language_model.layers — the transformer stack of the language model, and nothing else. Three groups were deliberately held back at 16-bit:

The vision tower is untouched — no entry in the quantization config refers to it. That is why image understanding survives the compression intact.

AutoRound ran with batch_size 4, gradient_accumulate_steps 2 and low_gpu_mem_usage true. The checkpoint does not record the calibration set, nsamples, seqlen or iters, so those are not stated here rather than guessed.

Measured on one Arc Pro B70

Numbers only mean something with their conditions attached, so each block below states them. They are not interchangeable — a single-user figure without reasoning and a three-user figure at xhigh describe different workloads.

Single user, reasoning off, XPU graphs on (the serving config below):

Decode speed does not depend on prompt size; waiting time does: wait ≈ prompt ÷ 1800, then ~32 tok/s.

Three concurrent users, reasoning on at `xhigh`, eager mode, twelve tasks per run — a deliberately heavy workload, and the reason these aggregates sit below the single-user figure above:

The shape is the point: throughput falls sharply as context grows, far more than it falls with added users. Keeping conversations compact is the most effective performance lever on this card. These three rows predate enabling XPU graphs; expect them to improve, but take the ratios rather than the absolutes.

Serving

docker run -d --name vllm --stop-timeout 120 \
  --device /dev/dri:/dev/dri -v /dev/dri/by-path:/dev/dri/by-path:ro \
  --group-add "$(getent group render | cut -d: -f3)" \
  --group-add "$(getent group video  | cut -d: -f3)" \
  --shm-size 2g \
  -v /srv/models:/models:ro -p 127.0.0.1:8000:8000 \
  -e VLLM_WORKER_MULTIPROC_METHOD=spawn \
  -e DISABLE_ESIMD_GDN_OUTPROJ=1 -e DISABLE_ESIMD_PAGE_ATTN=1 \
  intel/llm-scaler-vllm:0.21.0-b3.1 \
    --model /models/Qwen3.8-27B-int4-g128-sym-autoround \
    --served-model-name qwen \
    --dtype float16 --kv-cache-dtype fp8_e4m3 --max-model-len 131072 \
    --max-num-seqs 3 --max-num-batched-tokens 8192 --block-size 64 \
    --gpu-memory-utilization 0.95 --enable-prefix-caching \
    --trust-remote-code --limit-mm-per-prompt '{"image":3}' \
    --enable-auto-tool-choice --tool-call-parser qwen3_coder \
    --reasoning-parser qwen3

Host requirements

Why this configuration

Each of these was chosen against a measurement, and several run counter to what works on CUDA.

`--dtype float16` — set it explicitly rather than inheriting from config.json. It is the dtype this stack is built around on Battlemage, and the one to use for coherent output.

`--kv-cache-dtype fp8_e4m3` — both fast and roomy here. It yields 318,317 KV slots against 166,285 for fp16, nearly double, and stayed ahead of fp16 on throughput at every context size we compared. Use the explicit e4m3 name.

`DISABLE_ESIMD_GDN_OUTPROJ=1` — required for int4 weights. Intel's ESIMD fast path for the Gated-DeltaNet output projection is built for unquantized layers; switching it off puts the model on the general path, which is exactly where an int4 checkpoint belongs.

`DISABLE_ESIMD_PAGE_ATTN=1` — Intel's ESIMD paged attention is optimised for GQA ratios of 2, or ≥4 and divisible by 4. Qwen3.8-27B runs 24 query heads against 4 KV heads — a ratio of 6, which the generic xe_2 decode path handles directly and faster.

The gain grows with context. Switching ESIMD paged attention off improved aggregate throughput for three concurrent users by:

Which of those applies to you depends entirely on how long your conversations get — the longer they run, the more this flag is worth.

XPU graphs enabled (i.e. no --enforce-eager) — worth +29 %. This is the biggest single win and the most surprising one, since graphs are commonly avoided on Battlemage:

Cost: about 66 s of extra startup, once (62.8 s compile + 3 s capture), and a small amount of KV capacity — 318,317 slots in eager mode against 314,572 with graphs, i.e. roughly 1.2 %. vLLM separately reports 0.89 GiB taken during graph capture; that figure and the 3,745-slot difference are not the same accounting, so do not read one as the cause of the other. This was established against image intel/llm-scaler-vllm:0.21.0-b3.1, fp8_e4m3 and GuC firmware 70.72.1 — if you change any of those three, re-check before relying on it.

`--max-model-len 131072` rather than the native 262,144. The value also caps a single request: at 262,144 one user could occupy the entire cache and starve the other two. 131,072 allows 100k of context plus 31k of output and keeps multi-user service intact.

One observation we can report but not explain: measured KV capacity rose with larger --max-model-len on this stack — 293,187 slots at 65,536, 318,317 at 131,072, 332,570 at 262,144. That runs against the usual intuition that a longer maximum costs cache. Take it as a reason not to shrink the value hoping to buy capacity, and verify on your own build before relying on it.

`--max-num-seqs 3` — the design point for one B70. Idle sessions cost nothing; vLLM allocates KV only for requests actually in flight, so a single active writer gets the whole card. Beyond three concurrent generations, requests queue.

`--limit-mm-per-prompt '{"image":3}'` — set a limit explicitly. Without the flag there is no cap, and one request can carry an arbitrary number of images. Three covers multi-page document work; raise it if your workload needs more, keeping in mind that the vision tower runs unquantized and image tokens are correspondingly expensive.

Note this caps images only. The base model also accepts video, which is by far the most expensive input path here — add a video entry to the same flag (or set it to 0) rather than leaving it uncapped.

`--block-size 64` — the page size all measurements here were taken with. We did not compare block sizes, so treat this as the known-good value rather than a tuned one.

`--max-num-batched-tokens 8192` — matches the compile range vLLM captures graphs for (compile range (1, 8192)), so batched prefill stays inside the captured shapes.

`--gpu-memory-utilization 0.95` — the card has no display attached and nothing else competes for it; 0.95 turns the remaining headroom into KV cache. Leave lower if the GPU also drives a desktop.

`--enable-prefix-caching` — pays for itself in any chat workload, where each turn resends the whole conversation as prefix. Note it also means a repeated identical prompt measures the cache, not the model, so vary the leading bytes when benchmarking.

`/dev/dri/by-path` mount — required alongside --device, which passes only the device nodes. oneCCL scans that directory during initialisation.

`--stop-timeout 120` (or stop_grace_period: 120s in compose) — give the server room to finish in-flight work before it exits. Cutting a busy GPU container short can leave the card's scheduler in a state that only a reboot clears, so let it drain.

Speculative decoding (MTP): leave it off on this card. It is the largest win on CUDA and a loss here — 17–28 % slower above 32k context, because each user occupies triple query rows, KV slots drop by 24 %, and supports_spec_as_decode=False routes every step through prefill attention.

Which kernels this actually runs on

Worth knowing, because it explains the flags above and it is not what a CUDA user would expect.

A note on the name first. The sym in this repo's name describes the numeric scheme: symmetric int4, zero-point fixed at zero. vLLM separately has a loader class called sym_int4, and the two are not the same thing. Dispatch keys off the quant_method: auto-round string in the config, which vLLM maps to `quantization=inc` (Intel Neural Compressor) — so these weights are symmetric int4 loaded through the inc path, not through the class that shares the word.

That distinction has a consequence: several int4 ESIMD fast paths (_use_esimd_int4_gemv, _moe_int4_esimd_ok, _gdn_proj_int4_esimd_ok) test for the loader name sym_int4 and therefore stay inactive here, even though the underlying tensors are byte-identical to what that loader would produce. Both loaders end up calling the same int4_gemm_w4a16 kernel.

The two DISABLE_ESIMD_* environment switches are unrelated to that gate — they address two different places:

DISABLE_ESIMD_GDN_OUTPROJ=1 covers a spot that does not check quantization at all. The eligibility test for the Gated-DeltaNet output projection reaches straight for self.out_proj.weight, which on an int4 layer is qweight — so it has to be switched off explicitly rather than declining on its own.

DISABLE_ESIMD_PAGE_ATTN=1 is about attention shape, not quantization at all (see the GQA note below).

That leaves this arrangement on a B70:

The attention choice is the one with a measurable price tag. Intel's ESIMD paged attention is written for GQA ratios of 2, or ≥4 and divisible by 4 (csrc/eagle/eagle.sycl). This model runs 24 query heads against 4 KV heads — a ratio of 6, matching neither. The xe2 path handles ratio 6 directly, and does so faster; hence `DISABLEESIMDPAGEATTN=1`.

Getting the most out of it

Choose reasoning per request, and use the effort levels. Reasoning does not slow generation down — the token rate is essentially the same either way. It adds a second, invisible answer written before the visible one, and you pay for it in waiting time.

All figures below were taken at the default effort, xhigh. Measured on an 8,000-token prompt: the first word of the answer arrived after ~4.5 s with reasoning off, and after 77–100 s with it on. On a 500-token prompt: 0.3 s against 30–48 s. As a rule of thumb, roughly 40 seconds per 1,000 reasoning tokens at this decode rate.

Reasoning length is not predictable, even at temperature 0: the same short prompt produced traces of roughly 750 and 1,180 tokens on consecutive attempts. On long report-style tasks we have seen traces reach the tens of thousands of tokens.

Two consequences worth planning for:

• —Interactive chat is a different product with enable_thinking: false or reasoning_effort: low. Unattended work — overnight agents, document pipelines — is where xhigh earns its keep.
• —If you cap max_tokens with reasoning on, size the cap for the trace plus the answer. A budget smaller than the trace leaves nothing for the reply.

Keep working context compact. Look again at the throughput table: 29k of context per user gives nearly seven times the aggregate throughput of 100k. Summarising or rolling older turns out of the prompt buys more speed than any kernel flag.

Tool calling holds up as conversations grow. It behaves consistently at 20k and 60k of context, which is the case that actually matters in production — tools are rarely called on an empty prompt.

Plan concurrency around generations, not seats. Ten people with the tab open cost nothing. Three writing at once is the design point; the fourth waits.

Provenance and license

Quantized from Qwen/Qwen3.8-27B, released under Apache 2.0; this quantization carries the same license.

Throughput figures above come from a single Intel Arc Pro B70 running the configuration shown. Conditions are stated so you can reproduce them on your own card — and quality is best judged on your own workload.