melange-qcom

Melange Qualcomm

Use melange-qcom for the complete workflow. It shares Melange accounts, credentials, repositories, and local state with melange, but filters report and target presentation to the reviewed Qualcomm fleet.

Choose CLI or MCP

Prefer the connected Qualcomm MCP server when its tools are available. Start it with melange-qcom mcp; use the CLI for auth, local downloads, advanced uploads, or raw API access.

The api command is an intentionally unfiltered escape hatch. Never describe its report, target, device, or deployment data as Qualcomm-filtered. Apply and disclose filtering yourself when the dedicated command or MCP tool cannot do the job.

Authenticate and read structured output

melange-qcom auth login
melange-qcom auth status --json

Use MELANGE_API_KEY or MELANGE_API_KEY_FILE for headless environments. Pass --no-input in scripts. Branch on exit codes: 0 success, 1 API/operation error, 2 invocation error, 4 authentication error, and 130 interruption.

Use --json or --jq for agent work. Standard commands keep Melange's JSON contract. Qualcomm report and target commands return a re-marshaled filtered envelope with qualcomm_filter metadata; do not claim byte-exact API passthrough for those commands.

Keep identifiers distinct:

• ACCOUNT/REPO identifies a repository.
• MODEL_KEY identifies one converted model version.
• TARGET_ID identifies one downloadable artifact.

Resolve them in that order and never parse opaque model or target identifiers.

Manage and convert models

Check entitlement immediately before uploading:

melange-qcom plan --jq .plan
melange-qcom usage quotas --json

If .model_uploads.remaining == 0, stop. Attribute the restriction to the Free or Lite plan only after reading plan; otherwise call it an exhausted monthly quota.

Create a repository, validate the local manifest, then start conversion without blocking the user:

repo="$(melange-qcom repo create demo --private --jq .full_name)"
melange-qcom model upload -R "$repo" model.onnx --input sample.npy --dry-run --json
created="$(melange-qcom model upload -R "$repo" model.onnx --input sample.npy --json)"
model_key="$(printf '%s\n' "$created" | jq -er .model.key)"
melange-qcom model status "$model_key" -R "$repo" --json

For Hugging Face models, use:

melange-qcom model import ORG/MODEL -R "$repo" --json

Do not use --wait while a person is waiting. Report the current public state (converting, optimizing, ready, or failed) and monitor in the background. At optimizing, say artifacts are already downloadable while benchmarks finish. Never invent a percentage: public progress is null.

Render this phase panel after upload, import, and status checks:

╭─ Conversion Pipeline ───────────────────────────────╮
│   ✔ CONVERTING   →   ◐ OPTIMIZING   →   ○ READY     │
│     complete           in progress       pending    │
╰─────────────────────────────────────────────────────╯

Use ✔ for completed, ◐ for current, ○ for pending, and ✖ for failure. On failure, report .stage and .failure_code verbatim without guessing.

Report Qualcomm benchmarks

Resolve a ready model, then request its report:

model_key="$(melange-qcom model list -R "$repo" \
  --jq '.results | (map(select(.is_default and .state=="ready")) + map(select(.state=="ready")))[0].key // empty')"
melange-qcom report view "$model_key" -R "$repo" --json

The CLI matches exact reviewed (marketing_name, soc) pairs. It hides known non-Qualcomm and unclassified device records, recomputes summaries, and returns:

{
  "qualcomm_filter": {
    "strategy": "fixed-fleet-v1",
    "matched_devices": 1,
    "hidden_non_qualcomm_records": 10,
    "hidden_unclassified_records": 0
  }
}

Treat the matched set as the reviewed Qualcomm fleet, not every Qualcomm device in existence. If the command exits 1 with no Qualcomm measurements, do not fall back to general melange report output.

Read and fill the matching template before answering:

• General models: assets/report-general.md
• LLMs: assets/report-llm.md

Print the report in the reply. Preserve missing values as - in tables and N/A in cards. Never compare metrics from different accelerators as a speedup.

Select Qualcomm targets

melange-qcom model targets "$model_key" -R "$repo" --json

The command retains targets whose compatibility.soc_manufacturer is Qualcomm, retains compatibility-null device-unscoped artifacts used by LLM/universal targets, and hides explicitly other vendors. Inspect qualcomm_filter before claiming target coverage. Choose artifacts from precision, quantization, compatibility, accelerator classes, and size—never from an inferred engine.

Download only after the user authorizes the billable operation:

melange-qcom model download "$model_key" -R "$repo" \
  --target TARGET_ID --output ./models --yes

Generate deployment code

The Qualcomm edition supports android-kotlin, android-java, and flutter. It rejects ios-swift. Kotlin and auto are defaults.

melange-qcom deploy options --json
melange-qcom deploy guide "$model_key" -R "$repo" \
  --language android-kotlin --mode auto
melange-qcom deploy guide "$model_key" -R "$repo" \
  --language flutter --mode speed --json

Use the guide's exact SDK fields and callbacks. Keep YOUR_PERSONAL_KEY as the placeholder; never interpolate, print, or persist the active credential. General-model tensor construction may remain a TODO when shapes and preprocessing are model-specific.

Resume interrupted uploads

Prefer the resume command printed when exit 130 occurred. Otherwise locate the active session and resume by its opaque id:

session_id="$(melange-qcom model upload --sessions -R "$repo" \
  --jq '.results | map(select(.state=="CREATED" or .state=="UPLOADING")) | first | .id // empty')"
test -n "$session_id"
melange-qcom model upload --resume "$session_id" -R "$repo"

Run uploads with --dry-run first, never retry exit 2 or 4, and never present a metric or device classification the filtered response did not carry.