Utility tools

Purpose

Reference for auxiliary toolkit commands:

• qwen35-inspect for safetensors prefix/size inspection,
• qwen35-render-mermaid for Mermaid-to-PNG markdown rendering.

When to use

• You want to validate which weight groups exist in a model directory.
• You need Mermaid diagrams rendered for PDF/export workflows.

Command map

Command	Purpose	Docs section
qwen35-inspect	Inspect safetensors prefixes and cumulative sizes	qwen35-inspect
qwen35-render-mermaid	Render Mermaid blocks to PNG + markdown image refs	qwen35-render-mermaid

qwen35-inspect

Purpose

Read safetensors shard headers (without loading tensor payloads), then print unique top-level prefixes and cumulative sizes.

Syntax

qwen35-inspect <local_model_dir>

Mechanics

1. Read shard headers:
   - Parse safetensors metadata only (no tensor payload load).
2. Group keys:
   - Aggregate by top-level prefix.
3. Report:
   - Print cumulative size per prefix and total size.

Examples

qwen35-inspect ./Qwen3.5-0.8B-bnb-4bit
qwen35-inspect ./Qwen3.5-text-0.8B-bnb-4bit

Output examples

Example output (VLM checkpoint)

Found 1 shard(s)

PREFIX                                            SIZE
─────────────────────────────────────────────────────
model.language_model                           0.71 GB
model.visual                                   0.19 GB

TOTAL                                          0.90 GB

Interpretation: VLM BNB checkpoint contains both text and visual branches.

Stable fields: prefix naming and table layout. Variable fields: exact sizes and shard count.

Example output (text-only checkpoint)

Found 1 shard(s)

PREFIX                                            SIZE
─────────────────────────────────────────────────────
model.language_model                           0.71 GB

TOTAL                                          0.71 GB

Interpretation: text-only BNB checkpoint contains only language-model weights.

Stable fields: prefix naming semantics and table columns. Variable fields: sizes and shard count.

Edge cases / limitations

• Local paths only (HF repo ids are not downloaded automatically).
• Output reflects safetensors headers and key structure, not runtime inference behavior.

qwen35-render-mermaid

Purpose

Render Mermaid code blocks in markdown to PNG files and produce an export-friendly markdown file with image links.

Syntax

qwen35-render-mermaid --input <doc>.md [options]

Mechanics

1. Parse markdown:
   - Find fenced `mermaid` blocks.
2. Render diagrams:
   - Generate PNG files via Mermaid CLI.
3. Rewrite output markdown:
   - Replace mermaid blocks with image references.
4. Save artifacts:
   - Output markdown + diagrams directory.

Options

Flag	Default	Description
--input	required	Input markdown file
--output	<input_stem>_pdf.md	Output markdown file
--diagrams-dir	<output_dir>/diagrams/	Directory for rendered PNG files
--scale	2.0	PNG render scale (2.0 default, 3.0 for print)

Setup (one-time)

npm install -g @mermaid-js/mermaid-cli
npx puppeteer browsers install chrome-headless-shell

Examples

# Default output
qwen35-render-mermaid --input README.md

# Custom output file
qwen35-render-mermaid --input README.md --output README_pdf.md

# Higher image scale
qwen35-render-mermaid --input README.md --scale 3.0

# Batch usage example (BNB 4-bit VLM cards)
qwen35-render-mermaid --scale 3.0 --input ./model_cards/README_Qwen3.5-bnb-4bit.md --output ./Qwen3.5-0.8B-bnb-4bit/README.md
qwen35-render-mermaid --scale 3.0 --input ./model_cards/README_Qwen3.5-bnb-4bit.md --output ./Qwen3.5-2B-bnb-4bit/README.md
qwen35-render-mermaid --scale 3.0 --input ./model_cards/README_Qwen3.5-bnb-4bit.md --output ./Qwen3.5-4B-bnb-4bit/README.md
qwen35-render-mermaid --scale 3.0 --input ./model_cards/README_Qwen3.5-bnb-4bit.md --output ./Qwen3.5-9B-bnb-4bit/README.md

Edge cases / limitations

• Requires Mermaid CLI + headless browser runtime.
• Designed for markdown files containing fenced mermaid blocks.

Related

• Models
• Conversion pipeline
• Quickstart