Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
20
192
# Protocol: distinct prompts ~1500 tok in, 64 tok out (except noted); solo server on GPU; M-series 128GB
## oMLX (server as-shipped, max_concurrent_requests=8 default)
model=Qwen3.8-27B-8bit N= 1 in≈1500 out=64 via=127.0.0.1:8000 wall= 5.3s lat_p50= 5.3s lat_max= 5.3s agg= 12.0 out-tok/s ok=1 err=0
model=Qwen3.8-27B-8bit N= 2 in≈1500 out=64 via=127.0.0.1:8000 wall= 7.1s lat_p50= 7.1s lat_max= 7.1s agg= 17.9 out-tok/s ok=2 err=0
model=Qwen3.8-27B-8bit N= 4 in≈1500 out=64 via=127.0.0.1:8000 wall= 12.8s lat_p50= 12.8s lat_max= 12.8s agg= 20.0 out-tok/s ok=4 err=0
model=Qwen3.8-27B-8bit N= 8 in≈1500 out=64 via=127.0.0.1:8000 wall= 25.4s lat_p50= 25.2s lat_max= 25.4s agg= 20.1 out-tok/s ok=8 err=0
model=Qwen3.8-27B-8bit N=16 in≈1500 out=64 via=127.0.0.1:8000 wall= 51.8s lat_p50= 51.1s lat_max= 51.7s agg= 19.8 out-tok/s ok=16 err=0
## oMLX decode-heavy (256 out)
model=Qwen3.8-27B-8bit N=16 in≈1500 out=256 via=127.0.0.1:8000 wall= 114.8s lat_p50= 114.2s lat_max= 114.8s agg= 35.7 out-tok/s ok=16 err=0
## SGLang-MLX patched (max-running-requests=16, radix cache, no_buffer mamba)
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N= 1 in≈1500 out=64 via=127.0.0.1:30000 wall= 6.2s lat_p50= 6.2s lat_max= 6.2s agg= 10.3 out-tok/s ok=1 err=0
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N= 2 in≈1500 out=64 via=127.0.0.1:30000 wall= 8.1s lat_p50= 8.1s lat_max= 8.1s agg= 15.9 out-tok/s ok=2 err=0
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N= 4 in≈1500 out=64 via=127.0.0.1:30000 wall= 12.9s lat_p50= 12.9s lat_max= 12.9s agg= 19.8 out-tok/s ok=4 err=0
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N= 8 in≈1500 out=64 via=127.0.0.1:30000 wall= 23.7s lat_p50= 23.7s lat_max= 23.7s agg= 21.6 out-tok/s ok=8 err=0
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N=16 in≈1500 out=64 via=127.0.0.1:30000 wall= 48.7s lat_p50= 42.7s lat_max= 48.7s agg= 21.0 out-tok/s ok=16 err=0
## SGLang decode-heavy (256 out)
model=Qwen3.8-27B-8bit/snapshots/815b83c0df8ffd1d1b5244cf75fd6ef14fca9ef9/ N=16 in≈1500 out=256 via=127.0.0.1:30000 wall= 86.5s lat_p50= 86.5s lat_max= 86.5s agg= 47.4 out-tok/s ok=16 err=0
## Ollama 0.32.13 GGUF Q8_0 (private instance :11500, OLLAMA_NUM_PARALLEL=16, ctx 8192/slot)
model=qwen3.8:atomic-q8 N= 1 in≈1500 out=64 via=127.0.0.1:11500 wall= 9.4s lat_p50= 9.4s lat_max= 9.4s agg= 6.8 out-tok/s ok=1 err=0
model=qwen3.8:atomic-q8 N= 2 in≈1500 out=64 via=127.0.0.1:11500 wall= 8.9s lat_p50= 8.9s lat_max= 8.9s agg= 14.4 out-tok/s ok=2 err=0
model=qwen3.8:atomic-q8 N= 4 in≈1500 out=64 via=127.0.0.1:11500 wall= 17.9s lat_p50= 14.4s lat_max= 17.9s agg= 14.3 out-tok/s ok=4 err=0
model=qwen3.8:atomic-q8 N= 8 in≈1500 out=64 via=127.0.0.1:11500 wall= 37.0s lat_p50= 24.1s lat_max= 37.0s agg= 13.8 out-tok/s ok=8 err=0
model=qwen3.8:atomic-q8 N=16 in≈1500 out=64 via=127.0.0.1:11500 wall= 75.2s lat_p50= 40.5s lat_max= 75.2s agg= 13.6 out-tok/s ok=16 err=0
## Ollama decode-heavy (256 out)
model=qwen3.8:atomic-q8 N=16 in≈1500 out=256 via=127.0.0.1:11500 wall= 244.0s lat_p50= 131.6s lat_max= 244.0s agg= 16.8 out-tok/s ok=16 err=0
# Native MTP speculative decoding, Qwen3.8-27B on oMLX, M5 Max 128 GB — 2026-08-18
# Head: EigenLabs/Qwen3.8-27B-MTP-bf16 merged into mlx-community checkpoints (mtp/merge_mtp.py)
# Bench: mtp/bench_stream.py (101-tok prompt, 300 new tokens, reasoning_effort=low, single stream)
## bf16 A/B (same model, same server, load-toggle mtp_enabled)
mlx-community--Qwen3.8-27B-bf16 MTP off: 300 tok in 32.41s -> 9.3 tok/s (TTFT 1.13s)
mlx-community--Qwen3.8-27B-bf16 MTP on : 300 tok in 11.58s -> 25.9 tok/s (TTFT 1.19s) +178%
## all tiers, MTP on
mlx-community--Qwen3.8-27B-8bit MTP on : 300 tok in 8.35s -> 35.9 tok/s (head quantized 8-bit/gs64)
mlx-community--Qwen3.8-27B-4bit MTP on : 300 tok in 4.87s -> 61.6 tok/s (head kept bf16)
## in-context (Claude Code via qclaude, bf16, 51k-token context, live session): ~14.4 tok/s (was 4-6)
## draft acceptance seen in oMLX logs (Claude Code traffic): 75-99% ; 2.5-3.6 tokens/cycle (depth 3)
## Claude Code prompt-stability fixes (qclaude): --disallowedTools LSP,... ; CLAUDE_CODE_TOTAL_TOKENS_REMINDER=off ; reasoning_effort=medium
## fresh 3-request agentic run, --dense (8bit): 29 s wall; turn-1 prefill 6.5k tok, turns 2-3 primed 831/1020 tok (prefix reused)
## before (bf16, full tools, xhigh): 87 s for the same run; the user's long session hit 2-min re-prefills per turn (53k of 76k tokens)
## Ollama (MLX runner) reference, same machine
qwen3.8:27b-mlx (Ollama library, modelopt mixed-precision, ships mtp.* tensors, self-draft): 300 tok in 4.5s -> 66.7 tok/s
## Prior claim in daily-driver-20260818.txt "MTPLX MTP shows no gain" is superseded: that side-car was never imported
## into the checkpoint index (oMLX: "MTPLX side-car detected but not imported"), so MTP was inactive in that A/B.

Qwen3.8-27B concurrent serving on Apple Silicon — benchmarks, patches & recipes

Research artifacts from making 16 concurrent Qwen3.8-27B requests work on a single Apple Silicon Mac (M-series, 128 GB unified memory) — across three serving engines, including the patches that make SGLang's native MLX backend serve this model for the first time.

Code / full history: https://github.com/bluehawana/Qwen3.827B-SGLang-mpbm5max Run it with Ollama: ollama run bluehawana/qwen3.8-27b-q8 Fast single-stream (MTP speculative decoding, oMLX): bluehawana/Qwen3.8-27B-{4bit,8bit,bf16}-MTP-MLX — see MTP.md

Headline results

Solo GPU, distinct ~1.5k-token prompts, 64-token outputs (aggregate tok/s):

N concurrent oMLX (MLX 8-bit) SGLang-MLX patched Ollama 0.32 (GGUF Q8_0)
1 12.0 10.3 6.8
4 20.0 19.8 14.3
8 20.1 21.6 13.8
16 19.8 21.0 13.6
16 × 256-token outputs 35.7 47.4 16.8

Every cell: N/N requests succeeded, zero errors. A 27B dense model genuinely serves 16 concurrent clients on one Mac, on all three engines.

Key findings

  1. Pin in-model concurrency, queue the rest (--max-running-requests / max_concurrent_requests / OLLAMA_NUM_PARALLEL) — the one idea that ports from CUDA serving stacks to Metal.
  2. Short-output loads are prefill-bound and plateau by N=4. Long outputs are where continuous batching multiplies throughput (up to 4.5× the single-stream decode rate).
  3. SGLang's radix-cache scheduler wins decode-heavy (+33 % over oMLX at N=16); oMLX wins single-user latency; Ollama is the easiest and fully stable but 1.5–3× slower.
  4. GPU contention is the #1 false negative. With a second 30 GB model resident, the same hardware "collapses at N=6" and eventually hits a Metal OOM. Most "Macs can't do concurrent serving" reports are this artifact. One model server at a time.

Files

File Contents
BENCHMARKS.md Three-way ladder + full analysis
RESEARCH.md Community survey (vllm-mlx, mlx-lm batching, HN/X findings)
OLLAMA.md Run + concurrent-serve with Ollama
OMLX.md Run + concurrent-serve with oMLX (best for 1–4 users)
MTP.md Native MTP speculative decoding on oMLX: bf16 9→26 tok/s, 8-bit 36, 4-bit 62 (single stream) + Claude Code launcher fixes
mtp-20260818.txt Raw MTP benchmark output
ladder-20260816.txt Raw benchmark output
sglang-mlx-qwen38.patch 4-file patch: SGLang-MLX serves Qwen3.8-27B (hybrid-GDN cache, VLM checkpoint text-only, 2 batched-load device bugs)
PATCH-NOTES.md Why each patch exists, with crash signatures
bench_concurrency.py Stdlib-only benchmark harness

Reproduce the SGLang path

git clone https://github.com/bluehawana/Qwen3.827B-SGLang-mpbm5max
cd Qwen3.827B-SGLang-mpbm5max
./build-sglang-metal.sh          # needs Xcode Metal Toolchain; auto-applies patches
MAX_CONC=16 ./sglang-metal/serve-sglang-metal.sh

Key launch ingredients discovered here: SGLANG_USE_MLX=1, --json-model-override-args '{"language_model_only": true}' (runs the VLM checkpoint text-only — there is no CLI flag for this), --mamba-radix-cache-strategy no_buffer (hybrid-GDN default path is CUDA-only), plus the included patch.

Base model: Qwen/Qwen3.8-27B (Apache-2.0) · MLX quants: mlx-community · GGUF: AtomicChat

Downloads last month
117