club-3090/docs/engines

Inference engines for Qwen3.6-27B — comparison + quick recipes

This repo's main path is vLLM because it has the deepest support for Qwen3-Next features (vision, MTP, TurboQuant, full OpenAI API parity). But the model also runs on llama.cpp and SGLang with different trade-offs. This page compares the three; per-engine pages have setup instructions.

🔁 Coming from the README's Quick start? It already shipped you the vLLM path. Skim this comparison to see what the alternatives look like, then pick a per-engine page if you want to try one.

---

At a glance

Engine	Status on this stack	Per-stream TPS (1× 3090)	Max ctx (1× 3090)	Vision	Tool calls	Spec-decode	OpenAI API parity
vLLM ⭐	Validated, production-grade (this repo)	50-53 narr / 66-70 code	48K default · 75K IDE-agent · 198K vision · 214K text-only	✅	✅	✅ MTP n=3	✅ Full
llama.cpp	Works mainline + Luce DFlash fork for spec-decode	35-60 (varies by quant + KV type)	262K (Q4_K_M + q4_0 KV)	✅ (via mmproj)	⚠️ Limited (no auto-tool-choice in server)	✅ DFlash N=5 in fork	⚠️ Partial
SGLang	Blocked by same Marlin pad-sub-tile-n bug (vllm#40361 / sglang equivalent); EAGLE spec-decode separately blocked by GDN/DeltaNet rollback	n/a (untested at this state)	n/a	✅	✅	⚠️ EAGLE blocked on hybrid	✅ Full

---

Pros / cons matrix

vLLM ⭐

Pros:

• Deepest Qwen3-Next feature support upstream
• TurboQuant 3-bit KV cache (lets us reach 198K + vision or 214K text-only on a single 3090; 262K dual-card)
• MTP speculative decoding works out of the box
• Genesis patch ecosystem (Sandermage's tree fixes many compatibility edges)
• Full OpenAI API parity (chat, vision, tools, streaming, reasoning, structured output)
• Active development — bugs we hit get triaged within days

Cons:

• Heavyweight — Docker image is ~9 GB
• Longer cold-start (~2 min for compile + cudagraph capture)
• Sensitive to upstream API drift across nightly versions (we pin to a specific nightly SHA — 7a1eb8ac = 0.20.1rc1.dev16 since 2026-05-01 — to avoid this)
• Frontier features sometimes ship with bugs we have to patch around (the whole reason this repo exists)

When to pick: Production / serious local work / anything that needs the full feature set.

---

llama.cpp

Pros:

• Lightweight — single binary, ~50 MB
• Fastest cold-start (~30 sec)
• Lowest VRAM overhead (no inference framework taxes)
• GGUF support for many quant formats (Q4_K_M, Q5_K_S, IQ4_XS, etc.)
• Works on AMD + Intel + Apple Silicon (vLLM is NVIDIA-only)
• Active community, lots of distros / wrappers (Ollama, LM Studio, LocalAI, etc.)

Cons:

• Qwen3-Next family support is a moving target — needs the right binary build
• Server feature parity behind vLLM (no auto-tool-choice in upstream server; need wrapper)
• DFlash spec-decode requires a fork (Luce's llama-cpp-dflash)
• Concurrent serving is single-threaded by default (the server forks per request — sluggish under concurrent load)
• No TurboQuant equivalent → max usable context is much lower (~64K with Q4_K_M on 24 GB)

When to pick: Quick experiments, embedded use, non-NVIDIA hardware, when you want simplicity over feature completeness.

---

SGLang

Pros:

• Designed for high-throughput serving — RadixAttention prefix sharing, structured-output-aware scheduling
• Often beats vLLM by 10-30% on multi-tenant throughput when both work
• First-class OpenAI API
• Good support for batched structured output (constraint decoding)

Cons:

• Currently blocked on this stack by the same Marlin pad-sub-tile-n bug we hit on vLLM TP=2. Same kernel-line fix applies (would need a similar patch on SGLang's side or for them to pick up the upstream fix).
• EAGLE spec-decode (their MTP equivalent) is separately blocked by the DeltaNet/GDN hybrid layer not supporting KV rollback — this is a Qwen3-Next architectural issue, not SGLang-specific.
• Smaller community than vLLM; fewer eyes on Qwen3-Next bugs.

When to pick: Production multi-tenant serving on models that work cleanly on it (not yet Qwen3.6-27B-int4-AutoRound — track the unblock list below).

Watch list to unblock SGLang on this stack:

• Marlin pad-sub-tile-n landing (we filed PR #40361 on vLLM; the same fix applies to SGLang's Marlin call site)
• DeltaNet KV rollback support upstream (vllm#39931 / issue #40124 land would unblock EAGLE on Qwen3-Next family across engines)

---

How to choose

Your priority	Pick	Why
Full feature set, MTP spec-decode, OpenAI API parity	vLLM + Lorbus AutoRound	This repo's path. 51-70 TPS depending on workload, all features, prefill-safe at 48K default.
Maximum context (262K) on one 3090	llama.cpp + UD-Q3_K_XL or Q4_K_M + q4_0 KV	Smaller quants leave 8-10 GB headroom for KV at 262K. ~35-45 TPS sustained.
Best concurrent throughput on dual 3090	vLLM TP=2 + Turbo (TQ3)	4 streams at full 262K, ~200 TPS aggregate. See companion repo.
Non-NVIDIA hardware (AMD / Intel / Apple)	llama.cpp	Only engine with cross-platform support.
Lightest setup, fastest cold start	llama.cpp	Single binary, ~30s cold start. Good for embedded use, quick experiments.
High-throughput multi-tenant serving	SGLang (when unblocked — currently blocked on Qwen3.6)	RadixAttention prefix sharing wins at scale. Watch list in SGLANG.md.

---

Quant choice (orthogonal to engine choice)

The model itself comes in several quant formats. Engine-quant compatibility:

Quant	Disk size	Engine fit	Notes
AutoRound int4 (Lorbus)	~18-19 GB	vLLM ✅ · llama.cpp ❌ · SGLang (when unblocked)	This repo's choice. W4A16, group_size=128, BF16 mtp.fc head. Required for vLLM's MTP spec-decode.
GPTQ int4	~16.5-17 GB	vLLM ✅ · llama.cpp ❌ · SGLang ✅	Mature, broadly supported. Slightly smaller disk than AutoRound.
AWQ int4	~16-17 GB	vLLM ✅ · llama.cpp ❌ · SGLang ✅	Strong baseline, compatible with Marlin kernels.
GGUF Q4_K_M	~16.8 GB	llama.cpp ✅ · vLLM ⚠️ experimental · SGLang ❌	The default GGUF mid-range quant. Strong quality, broad ecosystem (Ollama, LM Studio, etc).
GGUF UD-Q3_K_XL (Unsloth)	~14.5 GB	llama.cpp ✅	Smaller than 4-bit options. Quality cost is small on Qwen3.6 (quantization-friendly), buys substantial KV cache room.
GGUF Q3_K_M	~13.6 GB	llama.cpp ✅	More aggressive 3-bit; quality cost real but acceptable for many workloads.

AutoRound vs GPTQ vs AWQ (within vLLM)

All three are 4-bit weight-only quantization for vLLM. Differences:

Aspect	AutoRound	GPTQ	AWQ
Method	Signed gradient descent jointly optimizing rounding + scaling	Layer-wise Hessian-based error minimization	Activation-aware salience scaling, then RTN
Calibration set	Small (~128-512 samples)	Larger (~1024-2048)	Small-medium
Quantization time	Minutes to ~1-2 hours for 27B	Slower for same model	Fast
Accuracy at 4-bit	Typically slightly best on hard reasoning (MMLU/GPQA/Math style)	Strong baseline; 0.5-2% behind AutoRound on average	Comparable to GPTQ; depends on tuning
Ultra-low bits (3, 2)	Strongest at <4 bit	Degrades faster below 4 bit	Middle of the pack
Marlin kernel support	✅ (via the kernel-line fix in our vllm#40361)	✅ (mature)	✅
Ecosystem	Newer, growing fast (Intel-maintained)	Most mature, broadest tool support	Strong vLLM/SGLang support

Why we picked AutoRound for this repo: Lorbus's AutoRound quant ships mtp.fc.weight as BF16 (preserved at higher precision), which lets vLLM's Qwen3_5MTP loader actually load the head and run multi-token prediction at high acceptance rates (~80% per-position-1, AL ~3.5). GPTQ-quantized variants of the MTP head silently fail to load → 0% draft acceptance. So AutoRound isn't just "slightly better quality" here — it's the only path to working MTP spec-decode in vLLM today.

If MTP isn't a priority for your workload, GPTQ or AWQ are equally valid.

---

Per-engine pages

• VLLM.md — current setup (what this repo ships). Brief recap + tuning levers.
• LLAMA_CPP.md — quick GGUF recipe, vision via mmproj, Luce DFlash fork pointer for spec-decode, gotchas around server feature parity.
• SGLANG.md — current blocked state, what would unblock, when to revisit. TBD recipe placeholder until either Marlin pad lands upstream or DeltaNet rollback lands.

---

See also

• docs/INTERNALS.md — why this repo picked vLLM specifically (the 9-probe forensics + upstream tracker)
• docs/SINGLE_CARD.md and docs/DUAL_CARD.md — workload-specific configs by hardware count
• LEARNINGS.md (parent stack) — why vLLM, why these patches