docs: add self-hosted B300 devnote

[lipikaramaswamy]

Summary

Adds a developer note for running NeMo Anonymizer fully self-hosted on one Brev B300. The post covers the local vLLM Qwen endpoint, local GLiNER reference server, Anonymizer Replace/Substitute configuration, warm-batch throughput and cost results, and reproducibility notes. It also adds a hero image and entity-density chart.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Refactoring

Testing

• make test passes locally
• make check passes locally (format + lint + typecheck + lock-check)
• Added/updated tests for changes

Docs-only change; no tests were added. make check exited 0; its advisory ty step emitted existing diagnostics and is ignored by the make target.

Documentation

• If docs changed: make docs-build passes locally

Related Issues

Depends on #203

[greptile-apps]

Greptile Summary

This PR adds a new developer note documenting how to run NeMo Anonymizer fully self-hosted on a single Brev B300 instance, covering the local vLLM Qwen endpoint, GLiNER reference server setup, Anonymizer Replace/Substitute configuration, and measured warm-batch throughput and cost results. It also adds a hero image and an entity-density distribution chart.

• New self-hosted-anonymizer-b300.md devnote with dataset prep code, service startup commands, full Anonymizer config, and throughput/cost tables for 100- and 250-record batches.
• Two PNG assets (hero image, entity-density chart) added to docs/devnotes/posts/assets/.
• mkdocs.yml nav extended with a new entry pointing to the new devnote.

Confidence Score: 3/5

The devnote content is generally solid, but two concrete defects need fixing before the page is reliably usable as a reproducibility guide.

The sidebar nav title and the page H1 use different words, which readers will notice immediately. More consequentially, the bash startup command omits VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel even though the surrounding prose explicitly states the successful run required this flag on the B300; someone copying the command verbatim could end up with a different kernel path than was tested.

Both mkdocs.yml (nav title) and docs/devnotes/posts/self-hosted-anonymizer-b300.md (bash startup command) need targeted fixes before merging.

Important Files Changed

Filename	Overview
docs/devnotes/posts/self-hosted-anonymizer-b300.md	New devnote covering the self-hosted B300 setup, throughput results, and cost breakdown; the bash launch command is missing the VLLM_DISABLED_KERNELS env var that the body text describes as having been set during the run, reducing reproducibility.
mkdocs.yml	Adds a nav entry for the new devnote, but the nav title ("Fully Self-Hosted") does not match the page H1 ("Fully Contained"), creating a visible inconsistency in the docs.
docs/devnotes/posts/assets/self-hosted-anonymizer-b300-hero.png	New hero image asset referenced correctly by the devnote; no issues.
docs/devnotes/posts/assets/self-hosted-anonymizer-b300-entity-density.png	New entity-density chart asset referenced correctly by the devnote; no issues.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Load dataset\nHugging Face TAB train split] --> B[Sort by char_len + idx\nTake head 100 / 250]
    B --> C[Write Parquet\ntab_train_compact_N.parquet]

    subgraph LocalStack["Local Stack on B300 (localhost)"]
        D[vLLM serve\nnvidia/Qwen3.6-35B-A3B-NVFP4\n:8001]
        E[serve_gliner.py\nnvidia/gliner-pii\n:9000]
    end

    C --> F[Anonymizer.run]
    D --> F
    E --> F

    subgraph Pipeline["Anonymizer Replace/Substitute Pipeline"]
        F --> G[GLiNER detection\n:9000]
        G --> H[LLM validation\nqwen-detect-1 + qwen-detect-2\n:8001]
        H --> I[LLM augmentation\nqwen-prod :8001]
        I --> J[Substitute-map generation\nqwen-prod :8001]
    end

    J --> K[Anonymized output\nParquet]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[Load dataset\nHugging Face TAB train split] --> B[Sort by char_len + idx\nTake head 100 / 250]
    B --> C[Write Parquet\ntab_train_compact_N.parquet]

    subgraph LocalStack["Local Stack on B300 (localhost)"]
        D[vLLM serve\nnvidia/Qwen3.6-35B-A3B-NVFP4\n:8001]
        E[serve_gliner.py\nnvidia/gliner-pii\n:9000]
    end

    C --> F[Anonymizer.run]
    D --> F
    E --> F

    subgraph Pipeline["Anonymizer Replace/Substitute Pipeline"]
        F --> G[GLiNER detection\n:9000]
        G --> H[LLM validation\nqwen-detect-1 + qwen-detect-2\n:8001]
        H --> I[LLM augmentation\nqwen-prod :8001]
        I --> J[Substitute-map generation\nqwen-prod :8001]
    end

    J --> K[Anonymized output\nParquet]

Loading

_{Reviews (1): Last reviewed commit: "docs: use public run config import" | Re-trigger Greptile}

[greptile-apps]

Nav title doesn't match page H1

The mkdocs.yml entry reads "Running NeMo Anonymizer Fully Self-Hosted on One B300" but the document's H1 is "Running NeMo Anonymizer Fully Contained on One B300". The word "Self-Hosted" vs "Contained" produces a visible mismatch: a reader sees one title in the sidebar and a different one when they land on the page. Given the file name (self-hosted-anonymizer-b300.md), the PR title, and the PR description all use "self-hosted", aligning the H1 to match the nav title (or vice-versa) would remove the inconsistency.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

VLLM_DISABLED_KERNELS missing from reproducible command

The "Blackwell serving note" at line 149 states the successful run "also set VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel to avoid the FlashInfer FP8 scaled-MM path." That variable does not appear in the bash snippet above, so a reader following the command verbatim on a B300 may hit the FlashInfer FP8 path and get the run configuration that was explicitly avoided. Adding the variable to the snippet (alongside the existing VLLM_USE_FLASHINFER_SAMPLER=0) keeps the presented command consistent with what was actually executed.