Block vLLM/SGLang serve on non-Linux with clear error

[alfredoclarifai]

Summary

• Add early platform check when serving models that use vLLM or SGLang toolkits
• On macOS/Windows, these engines crash deep in C extensions with opaque AttributeError tracebacks
• Now fails fast with a clear message: what's wrong and what to do instead (cloud deploy or Ollama)
• Applied to both serve paths (API-connected and --grpc)

Test plan

• On macOS, run clarifai model serve . in a vLLM model directory — should get clear error instead of C extension traceback
• On macOS, run clarifai model serve --grpc in an SGLang model directory — same clear error
• On Linux, verify serve still works normally (platform check passes)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds an early OS/platform guard in the CLI model serving paths to fail fast (with a clearer UserError) when attempting to serve vLLM/SGLang-based models on non-Linux platforms.

Changes:

• Add non-Linux detection for vLLM/SGLang during clarifai model serve (API-connected) validation.
• Add similar detection during clarifai model serve --grpc validation.
• Reuse toolkit.provider (when present) alongside requirements.txt inspection to detect the engine.

Comments suppressed due to low confidence (1)

clarifai/cli/model.py:1404

• serve_cmd introduces toolkit_provider = config.get('toolkit', {}).get('provider'), but later in the same validation block the LM Studio branch still re-reads config.get('toolkit', {}).get('provider') instead of using toolkit_provider. Using the cached variable consistently would avoid duplicate lookups and keep the checks uniform (and reduces the chance of future drift between branches).

    toolkit_provider = config.get('toolkit', {}).get('provider')
    if _platform.system() != "Linux":
        for engine in ('vllm', 'sglang'):
            if engine in dependencies or toolkit_provider == engine:
                raise UserError(
                    f"{engine} is not supported on {_platform.system()}. It requires a Linux environment with GPU access.\n"
                    "  Use 'clarifai model deploy .' to run on cloud GPU, or switch to the Ollama or LM Studio toolkit for local serving."
                )

    if "ollama" in dependencies or toolkit_provider == 'ollama':

[Copilot]

This adds new platform-specific behavior (raising UserError on non-Linux when vLLM/SGLang is detected), but the existing CLI tests for clarifai model serve don’t appear to cover these branches. Adding unit tests that mock platform.system() (and set up requirements.txt/toolkit.provider for vllm/sglang) would prevent regressions and should cover both API-connected serve and the --grpc path.

[Copilot]

In _run_local_grpc, the non-Linux guard for vLLM/SGLang only runs inside if mode not in ("container", "env"). That means clarifai model serve --grpc --mode env|container on macOS/Windows will skip this early check and can still hit the same deep C-extension failures this PR is trying to avoid. Consider moving the platform/toolkit check outside the mode gate (using toolkit_provider and/or parsing requirements regardless of mode) so the behavior is consistent across all --grpc modes.

[luv-bansal]

@alfredoclarifai I have included these improvements in my branch, so closing this PR