vh
5e47a59b32
The multi-turn Gemma-4 CoT problem is solved. One MeroMero-v2 seat, one weight set,
two aliases: char-rp (prose) + char-rp-reasoning (streaming chain-of-thought).
The winning stack, traced from vLLM source by the four-arm brokkr/dwarf panel:
- vllm/vllm-openai:v0.26.0 — ships transformers 5.14.1 natively, below the
head_dim guard, so Gemma-4-31B loads with no pin and no custom image. It also
carries the #48217 streaming pre-arm fix.
- A patched chat template whose enable_thinking:true branch force-opens a BARE
<|channel> (not <|channel>thought\n -- full-open defeats _preprocess_feed's
injection). --chat-template override, no re-quant.
- Two served-names char-rp / char-rp-thinking; --reasoning-parser gemma4;
default enable_thinking:false. LiteLLM char-rp -> prose, char-rp-reasoning ->
the thinking served-name with enable_thinking:true.
Verified: streaming CoT split 6/6 direct on :8016 and 3/3 through the gateway;
char-rp prose clean on both transports with no trailing-token leak.
Two hard-won facts recorded in persistent-memory:
- STREAMING ONLY. Non-streaming can't split -- extract_reasoning never receives
prompt_token_ids so the pre-arm can't fire (a vLLM one-shot bug unchanged
across v0.24-0.27). Fine here: Lobe/OWUI stream. Upstream PR #49797 fixes
non-streaming too, landing ~v0.28.0 -- then it's a clean image bump.
- KEY-NAME TRAP: vLLM streams reasoning in delta.reasoning; LiteLLM normalizes
to delta.reasoning_content. I lost two false-negative test rounds to this.
Canonical: stacks/meromero-charrp/ (compose + patched_chat_template.jinja) and
stacks/litellm/conf/config.yaml. Rollback is the .env image line + dropping
--chat-template.
99 lines
4.0 KiB
YAML
99 lines
4.0 KiB
YAML
# meromero-charrp — G4-MeroMero-v2-31B NVFP4A16 char-rp PROSE seat (non-thinking, MULTIMODAL)
|
|
# on ana-ml2 GPU0. Replaces the retired Magidonia GGUF seat (char-rp-gguf / llama-charrp).
|
|
#
|
|
# Gemma-4 dense but with SLIDING-WINDOW attention -> KV-efficient, serves the full native 256K
|
|
# at ~2x concurrency. Vision enabled: preprocessor_config.json was materialized from
|
|
# processor_config.json's image_processor section (Gemma4ImageProcessor); audio is
|
|
# config-declared but weightless (no audio tensors) so it serves image + text.
|
|
#
|
|
# Co-located on GPU0 with vllm-aeon-gen (gen). util 0.52 leaves ~5GB GPU0 headroom
|
|
# (0.55 left only ~1.8GB). Tunables in .env.
|
|
|
|
name: meromero-charrp
|
|
|
|
services:
|
|
vllm-meromero-rp:
|
|
image: ${MEROMERO_IMAGE:-vllm/vllm-openai:latest}
|
|
container_name: ${MEROMERO_CONTAINER:-vllm-meromero-rp}
|
|
restart: unless-stopped
|
|
ipc: host
|
|
ports:
|
|
- "${MEROMERO_PORT:-8016}:8000"
|
|
volumes:
|
|
- /tank/aimodels:/tank/aimodels
|
|
environment:
|
|
- PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
|
|
- VLLM_API_KEY=${API_KEY:-}
|
|
command:
|
|
- ${MEROMERO_MODEL:-/tank/aimodels/meromero-v2-nvfp4-work/G4-MeroMero-v2-31B-NVFP4A16}
|
|
- --quantization
|
|
- compressed-tensors
|
|
- --served-model-name
|
|
- char-rp
|
|
- char-rp-thinking
|
|
# Tool-calling: Gemma-4 emits its OWN native syntax
|
|
# (<|tool_call>call:name{...}<tool_call|>), NOT the qwen3_coder XML the
|
|
# other seats use. vLLM 0.24 ships a matching `gemma4` parser whose token
|
|
# constants line up with this tokenizer's etc/eoc/escape tokens exactly.
|
|
# Without these two flags any tools-bearing request 400s outright.
|
|
- --tool-call-parser
|
|
- gemma4
|
|
- --enable-auto-tool-choice
|
|
# The gemma4 REASONING parser is what absorbs the <|channel>...<channel|>
|
|
# thought markers; without it they leak into `content` verbatim on the
|
|
# post-tool-response turn (upstream vllm #45834 — the chat template leaves
|
|
# the prompt inside an open channel block).
|
|
- --reasoning-parser
|
|
- gemma4
|
|
# Dwarf-panel CoT patch 2026-08-21: force-open <|channel> on enable_thinking:true
|
|
- --chat-template
|
|
- /tank/aimodels/meromero-v2-nvfp4-work/patched_chat_template.jinja
|
|
# MANDATORY companion to the reasoning parser on this seat. The parser
|
|
# reads enable_thinking from chat_template_kwargs and DEFAULTS IT TO TRUE
|
|
# (vllm/parser/gemma4.py:439). True makes is_reasoning_end() return False
|
|
# at a new turn, which pre-initialises the engine to REASONING -> ALL plain
|
|
# RP prose lands in reasoning_content with a NULL content, breaking every
|
|
# char-rp consumer. This template already defaults enable_thinking to false
|
|
# (chat_template.jinja:350), so passing it explicitly renders a BYTE-
|
|
# IDENTICAL prompt (verified across plain/tools/post-tool/system shapes) --
|
|
# it only corrects the parser's state machine. Do not remove.
|
|
- --default-chat-template-kwargs
|
|
- '{"enable_thinking": false}'
|
|
- --max-model-len
|
|
- "${MEROMERO_MAX_MODEL_LEN:-262144}"
|
|
- --max-num-seqs
|
|
- "${MEROMERO_MAX_NUM_SEQS:-32}"
|
|
- --gpu-memory-utilization
|
|
- "${MEROMERO_GPU_MEM_UTIL:-0.52}"
|
|
- --kv-cache-dtype
|
|
- fp8
|
|
- --trust-remote-code
|
|
deploy:
|
|
resources:
|
|
reservations:
|
|
devices:
|
|
- driver: nvidia
|
|
device_ids:
|
|
- "${MEROMERO_GPU_ID:-0}"
|
|
capabilities:
|
|
- gpu
|
|
healthcheck:
|
|
test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
|
|
interval: 30s
|
|
timeout: 10s
|
|
retries: 3
|
|
start_period: 600s
|
|
networks:
|
|
- tnet
|
|
labels:
|
|
- homepage.group=AI - Inference
|
|
- homepage.name=char-rp (MeroMero-v2 NVFP4, multimodal)
|
|
- homepage.icon=mdi-drama-masks
|
|
- homepage.description=G4-MeroMero-v2-31B NVFP4A16 non-thinking prose seat, vision-enabled, 256K (ana-ml2 GPU0)
|
|
- homepage.href=http://10.250.50.54:${MEROMERO_PORT:-8016}/docs
|
|
|
|
networks:
|
|
tnet:
|
|
name: traefik-net
|
|
external: true
|