Add remote exec capability for foundation models missing it

[hansent]

Manual testing results

I tested the workflows blocks with added remote exec functionality on a dev GPU with following results:

working:
gaze - works both with local and remote exec
depth estimation: works both with local and remote exec
SAM2: works both with local and remote exec
florence - works both with local and remote exec (some workflow block config causes broken results, but same for local/remote)
Moondream2 - works both with local and remote exec

not working/couldnt test:

• smolvlm: doesnt work local or remote because model download fails (currently disabled on serverless anyway)
• qwen: doesnt work local or remote because model download fails (currently disabled on serverless anyway)
• SAM3 3d: couldnt install dependencies get model working (disabled on serverless anyway)

I think its ok to add the missing remote exec functionality and endpoints handling for these already also.

1. HTTP Client Methods Added (inference_sdk/http/client.py)

• infer_lmm() - Generic LMM endpoint for Florence2, Moondream2, SmolVLM, Qwen models
• depth_estimation() - Depth estimation endpoint
• sam2_segment_image() - SAM2 segmentation endpoint
• sam3_3d_infer() - SAM3 3D object generation endpoint
• Added async variants for all methods

2. HTTP API Endpoint Added (inference/core/interfaces/http/http_api.py)

• /sam3_3d/infer - New endpoint for SAM3 3D object generation with JSON-serializable response (base64-encoded binary data)

3. Workflow Blocks Updated with Remote Execution

All blocks now support StepExecutionMode.REMOTE:

Block	Client Method Used
Gaze v1	detect_gazes()
Depth Estimation v1	depth_estimation()
Segment Anything 2 v1	sam2_segment_image()
Florence2 v1	infer_lmm()
Moondream2 v1	infer_lmm()
SmolVLM v1	infer_lmm()
Qwen2.5-VL v1	infer_lmm()
Qwen3-VL v1	infer_lmm()
Segment Anything 3 3D v1	sam3_3d_infer()

4. Unit Tests Added

• test_gaze_remote.py - Gaze detection remote execution tests
• test_depth_estimation.py - Depth estimation tests
• test_segment_anything2.py - SAM2 tests
• test_vlm_remote_execution.py - Tests for Florence2, Moondream2, SmolVLM, Qwen2.5-VL, Qwen3-VL
• test_segment_anything3_3d.py - SAM3 3D tests

5. Documentation Updated

Added "Execution Modes" sections to:

• docs/foundation/florence2.md
• docs/foundation/gaze.md
• docs/foundation/depth_estimation.md
• docs/foundation/sam2.md
• docs/foundation/moondream2.md
• docs/foundation/smolvlm.md
• docs/foundation/sam3_3d.md

[codeflash-ai]

⚡️ Codeflash found optimizations for this PR

📄 45% (0.45x) speedup for Qwen3VLBlockV1.run in inference/core/workflows/core_steps/models/foundation/qwen3vl/v1.py

⏱️ Runtime : 5.64 milliseconds → 3.90 milliseconds (best of 71 runs)

A dependent PR with the suggested changes has been created. Please review:

• ⚡️ Speed up method Qwen3VLBlockV1.run by 45% in PR #1968 (remote-exec-for-all-models) #1971

If you approve, it will be merged into this PR (branch remote-exec-for-all-models).

[bigbitbus]

LGTM.

[grzegorz-roboflow]

LGTM