You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs/usermanuals/en/deployment/deploy-targets.md: false claim — The 'What's auto-detected' table covers chat template, stop tokens, max_model_len from model.max_length, quantisation and tensor parall… (DOC-D3B08-010) #260
The document says. The 'What's auto-detected' table covers chat template, stop tokens, max_model_len from model.max_length, quantisation and tensor parallelism; 'You can override any auto-detection with explicit YAML'.
What the code does. generate_deploy_config inspects nothing: no tokenizer, config, GPU or training YAML (deploy.py:218-330). max length comes from --max-length (default 4096). Deploy takes no YAML, as line 127 of the same page says. The flag table (lines 139-142) also omits the real knobs: --system, --max-length, --gpu-memory-utilization, --port, --trust-remote-code, --vendor and --output-format.
User impact. Users rely on detection that never happens and miss the flags that do exist.
Proposed correction. Delete the table and the YAML sentence. List all deploy flags with their defaults (--max-length 4096, --gpu-memory-utilization 0.90, --port 8080, --vendor aws, --system, --trust-remote-code).
Turkish mirror. the Turkish page has the same error; fix both in the same PR
Acceptance criteria
Each statement is re-checked against the code at fix time.
Documentation listed above is corrected in English and in its Turkish mirror in the same PR.
Every finding in this issue is resolved or explicitly closed with a reason.
Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.
Activity