fix(config): Qwen3.6-35B-A3B bleibt warmes Gehirn (90 t/s MoE), Qwen3.8-27B on-demand bereitgestellt
This commit is contained in:
@@ -3,27 +3,32 @@ healthCheckTimeout: 300
|
|||||||
globalTTL: 0
|
globalTTL: 0
|
||||||
|
|
||||||
models:
|
models:
|
||||||
Qwen3.8-27B:
|
Qwen3.6-35B-A3B:
|
||||||
# Qwen 3.8 27B: Dichtes 27B-Modell mit hybrider Linear-Attention (48/64 Schichten linear),
|
# parallel 2 + c 131072: 2 Slots à 65k (Slot 2 = Mem0-Lern-Extraktion, blockiert Voice-Turns
|
||||||
# nativem Multimodal-Support (mmproj-BF16), Tool-Calling via --jinja.
|
# nicht mehr); KV Q8_0 macht das speicherneutral zu vorher (65k f16). Bench 2026-07-02:
|
||||||
# parallel 2 + c 131072: 2 Slots à 65k (Slot 2 = Mem0-Lern-Extraktion); KV Q8_0 für hohe Qualität.
|
# Q8_0 kostet 0 t/s, MTP n-max 3 = +26% vs. ohne Spec. (Q4_K-KV stand mal im Template,
|
||||||
|
# lief aber NIE live — bewusst bei Q8_0 geblieben: Qualität vor ein paar GB, seit der
|
||||||
|
# Warm-Set-Diät 07.07. ist RAM nicht mehr der Engpass.)
|
||||||
cmd: |
|
cmd: |
|
||||||
llama-server -m /srv/models/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --mmproj /srv/models/Qwen3.8-27B-GGUF/mmproj-BF16.gguf --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0
|
llama-server -m /srv/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0 --spec-type draft-dflash --spec-draft-model /srv/models/Qwen3.6-35B-A3B-DFlash-GGUF/giocom-Qwen3.6-35B-A3B-DFlash-Q8_0.gguf
|
||||||
ttl: 0
|
ttl: 0
|
||||||
aliases:
|
aliases:
|
||||||
- hermes
|
- hermes
|
||||||
- fast
|
- fast
|
||||||
|
# context = nutzbarer Kontext PRO REQUEST (c geteilt durch parallel-Slots).
|
||||||
capabilities:
|
capabilities:
|
||||||
in: [text, image]
|
in: [text]
|
||||||
out: [text]
|
out: [text]
|
||||||
tools: true
|
tools: true
|
||||||
context: 65536
|
context: 65536
|
||||||
Qwen3.6-35B-A3B:
|
Qwen3.8-27B:
|
||||||
|
# Qwen 3.8 27B: Dichtes 27B-Modell mit hybrider Linear-Attention (48/64 Schichten linear),
|
||||||
|
# nativem Multimodal-Support (mmproj-BF16), Tool-Calling via --jinja.
|
||||||
cmd: |
|
cmd: |
|
||||||
llama-server -m /srv/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0 --spec-type draft-dflash --spec-draft-model /srv/models/Qwen3.6-35B-A3B-DFlash-GGUF/giocom-Qwen3.6-35B-A3B-DFlash-Q8_0.gguf
|
llama-server -m /srv/models/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --mmproj /srv/models/Qwen3.8-27B-GGUF/mmproj-BF16.gguf --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0
|
||||||
ttl: 300
|
ttl: 300
|
||||||
capabilities:
|
capabilities:
|
||||||
in: [text]
|
in: [text, image]
|
||||||
out: [text]
|
out: [text]
|
||||||
tools: true
|
tools: true
|
||||||
context: 65536
|
context: 65536
|
||||||
@@ -118,4 +123,4 @@ groups:
|
|||||||
members:
|
members:
|
||||||
- Qwen3-Embedding-0.6B
|
- Qwen3-Embedding-0.6B
|
||||||
- Qwen3-Reranker-0.6B
|
- Qwen3-Reranker-0.6B
|
||||||
- Qwen3.8-27B
|
- Qwen3.6-35B-A3B
|
||||||
|
|||||||
Reference in New Issue
Block a user