Recipe

qwen3-6-27b-mq4-radeon-ai-pro-r9700-32gb-hipfire-tp1

qwen3-6-27b-mq4-radeon-ai-pro-r9700-32gb-hipfire-tp1

JSON API Find compatible recipes

Observed LocalMaxxing leaderboard run. Evidence for compatibility, not an executable launch contract.

Record

Status: candidate
Source: localmaxxing
Engine: hipfire
Engine version: 0.3.0+bcd1ef0
Accelerators: 1
Tensor parallel: 1
Context tokens: 2,048
chat: unknown
reasoning: unknown
tools: unknown
vision: unknown

Hugging Face model card

Identity

https://huggingface.co/Qwen/Qwen3.6-27B

Repository: Qwen/Qwen3.6-27B
Status: known
Link type: Exact Hub repository

Public Hugging Face repository confirmed by the Hub API.

Candidate: useful compatibility or speed evidence. The registry does not offer Run until promotion requirements are met.

Observed configuration

hipfire

Evidence only · candidate · reference

Candidate evidence — not a Run contract

Source: https://www.localmaxxing.com/en/runs/cms8cr7d700efpf01ghl7scrj

Observed source tokens

Mechanical split of the source command. Unverified against the engine CLI. The registry does not offer Run for this recipe.

./target/release/examples/dflash_spec_demo
--target
~/.hipfire/models/qwen3.6-27b-awq.mq4
--draft
~/.hipfire/models/qwen36-27b-dflash-mq4.hf4
--prompt-file
benchmarks/prompts/merge_sort_thinking_off.txt
--max
256
--temp
0.0
--no-chatml
--kv-mode
q8
--ctx
2048

Flag	Value
`--target`	`~/.hipfire/models/qwen3.6-27b-awq.mq4`
`--draft`	`~/.hipfire/models/qwen36-27b-dflash-mq4.hf4`
`--prompt-file`	`benchmarks/prompts/merge_sort_thinking_off.txt`
`--max`	`256`
`--temp`	`0.0`
`--kv-mode`	`q8`
`--ctx`	`2048`

Environment

Variable	Value
`HIPFIRE_HFQ4G256_LDSSTAGE`	`1`
`HIPFIRE_Q8_PREFILL_M4`	`0`
`HIP_VISIBLE_DEVICES`	`3`

Measured speed

Concurrency	Context	Prefill	Decode	TTFT ms	Status	Sweep
1	2,048	466.6	278	103.1	observed	qwen3-6-27b-mq4-radeon-ai-pro-r9700-32gb-hipfire-tp1-sweep

Remaining fields

Identity, launch, related records, and measured speed are shown above. This is the rest of the normalized record.

hardware count: 1
hardware id: radeon-ai-pro-r9700-32gb
id: qwen3-6-27b-mq4-radeon-ai-pro-r9700-32gb-hipfire-tp1
model instance id: qwen-qwen3-6-27b--mq4
recipe source: localmaxxing
schema version: local-ai-registry/v1
speed sweep ids: qwen3-6-27b-mq4-radeon-ai-pro-r9700-32gb-hipfire-tp1-sweep
status: candidate

capabilities

engine

name: hipfire
version: 0.3.0+bcd1ef0

serving

max context tokens: 2,048
tensor parallel: 1

Provenance & metadata (3)

facts

capabilities.chat · provenance · captured at: 2026-08-30T09:10:02Z

capabilities.chat · reason capability-not-verified

capabilities.chat · state unknown

capabilities.reasoning · provenance · captured at: 2026-08-30T09:10:02Z

capabilities.reasoning · reason capability-not-verified

capabilities.reasoning · state unknown

capabilities.tools · provenance · captured at: 2026-08-30T09:10:02Z

capabilities.tools · reason capability-not-verified

capabilities.tools · state unknown

capabilities.vision · provenance · captured at: 2026-08-30T09:10:02Z

capabilities.vision · reason capability-not-verified

capabilities.vision · state unknown

engine.graph mode · provenance · captured at: 2026-08-30T09:10:02Z

engine.graph mode · reason runtime-detail-not-published

engine.graph mode · state unknown

serving.kv cache tokens · provenance · captured at: 2026-08-30T09:10:02Z

serving.kv cache tokens · reason kv-cache-capacity-not-published

serving.kv cache tokens · state unknown

serving.max concurrency · provenance · captured at: 2026-08-31T23:03:15Z

serving.max concurrency · reason server-capacity-not-evidenced

serving.max concurrency · state unknown

metadata

localmaxxing · backend: rocm
localmaxxing · batch size: 1
localmaxxing · hardware label: Radeon AI Pro R9700
localmaxxing · notes: Single client, single GPU, non-batched. NOT tensor-parallel, NOT pipeline-parallel, NOT a batched serve: one request at a time via the standalone dflash_spec_demo binary, not an HTTP server. Host has 4x R9700; this uses exactly one (HIP dev 3). Pure DFlash speculative decode: 5-layer trained draft head, block B=16, adaptive-B 8..16 which settled at B=16 for 100% of cycles. No MTP, n-gram, prompt-lookup, tree, or CASK. Greedy temp 0.0, deterministic. Identical every rep: tau=11.3846, accept_rate=0.7590, cycles=13, committed=174, accepted=148, emitted=162, EOS reached. 3 fresh-process reps after one discarded JIT warmup, GPU-lock serialized, clocks untouched: decode 278.24/277.97/276.92 -> median 277.97 (spread 0.47%); prefill 466.57/466.68/461.68; ttft 102.97/103.13/103.86. Two non-default flags. HIPFIRE_HFQ4G256_LDSSTAGE=1 (gfx12 LDS-staged HFQ4-G256 WMMA, default off): same-binary A/B gave 254.88 with it off, i.e. +9.06%. HIPFIRE_Q8_PREFILL_M4=0 disables the four-query Q8 prefill tile, a ~4% LOSS on the speculative path though a win on AR prefill. Routing verified not assumed: LDS-on must emit gemm_gate_up_hfq4g256_wmma_gfx12_ldsstage.hsaco (17448 B) and gemm_hfq4g256_residual_wmma_gfx12_ldsstage.hsaco (15400 B); both present. NOT bit-exact by construction: LDS staging reorders f32 K-accumulation across 8 waves. Output matched byte-for-byte (md5 4d3a55e4) but that is one prompt at temp 0, not a quality result. Quantized artifact: schuttdev/hipfire-qwen3.6-27b @ f9b326a657f1 (qwen3.6-27b.mq4). Provenance: tree bcd1ef0082e7f7f7a1eff9252da5d76e17d042b0 (0 dirty); binary sha256 a107cf8640abf1d7; trunk sha256 86a5f80fd29d545a; draft md5 204c4c4c; prompt md5 253c7ac5; ROCm/HIP 7.14.60850. peakVramGb = 17034 MiB reported by the runtime. Unedited logs retained. Coherence eyeballed (correct Python merge_sort, ends <|im_end|> then <|endoftext|>), not a scored gate. No AR baseline rerun.
localmaxxing · observed command: HIP_VISIBLE_DEVICES=3 HIPFIRE_HFQ4G256_LDSSTAGE=1 HIPFIRE_Q8_PREFILL_M4=0 ./target/release/examples/dflash_spec_demo --target ~/.hipfire/models/qwen3.6-27b-awq.mq4 --draft ~/.hipfire/models/qwen36-27b-dflash-mq4.hf4 --prompt-file benchmarks/prompts/merge_sort_thinking_off.txt --max 256 --temp 0.0 --no-chatml --kv-mode q8 --ctx 2048
localmaxxing · run id: cms8cr7d700efpf01ghl7scrj
localmaxxing · tokenized · arguments: ./target/release/examples/dflash_spec_demo, --target, ~/.hipfire/models/qwen3.6-27b-awq.mq4, --draft, ~/.hipfire/models/qwen36-27b-dflash-mq4.hf4, --prompt-file, benchmarks/prompts/merge_sort_thinking_off.txt, --max, 256, --temp, 0.0, --no-chatml, --kv-mode, q8, --ctx, 2048
localmaxxing · tokenized · environment · HIPFIRE HFQ4G256 LDSSTAGE: 1
localmaxxing · tokenized · environment · HIPFIRE Q8 PREFILL M4: 0
localmaxxing · tokenized · environment · HIP VISIBLE DEVICES: 3
localmaxxing · tokenized · fidelity: faithful

provenance

captured at: 2026-08-30T09:10:02Z

sources

captured at	kind	url
2026-08-30T09:10:02Z	normalized-recipe	www.localmaxxing.com/en/runs/cms8cr7d700efpf01ghl7scrj ↗