Spaces:

finerobotics
/

annotation-studio

Running

Fix skipped videos without a runtime limit; calibrate video pricing (#2)

578dc0f 24 days ago

11.4 kB

"""Curated, vision-tested HF routes. Server-owned pricing and request parameters.
This is a practical selection, not a universal quality ranking. Keep checked_at
and smoke tests current when changing any route. Never infer vision support from
model family names or silently substitute a similarly named checkpoint.
"""

from copy import deepcopy

DEFAULT_MODEL = "qwen38-27b"
MODELS = {
    "qwen38-27b": {
        "key": "qwen38-27b",
        "name": "Qwen3.8 27B",
        "model_id": "Qwen/Qwen3.8-27B",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "Qwen/Qwen3.8-27B:novita",
        "input_rate": 0.42,
        "output_rate": 3,
        "cached_input_rate": 0.085,
        "context_length": 1000000,
        "concurrency": 32,
        "generation_options": {"chat_template_kwargs": {"enable_thinking": False}},
        "featured": True,
        "description": "Current default · tested on 33 robot episodes",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "32 simultaneous vision requests tested",
        "model_url": "https://huggingface.co/Qwen/Qwen3.8-27B",
        "pricing_url": "https://novita.ai/pricing",
    },
    "qwen36-27b": {
        "key": "qwen36-27b",
        "name": "Qwen3.6 27B",
        "model_id": "Qwen/Qwen3.6-27B",
        "provider": "ovhcloud",
        "provider_name": "OVHcloud",
        "route": "Qwen/Qwen3.6-27B:ovhcloud",
        "input_rate": 0.47,
        "output_rate": 3.19,
        "cached_input_rate": 0.47,
        "context_length": 262144,
        "concurrency": 8,
        "generation_options": {"reasoning_effort": "none"},
        "featured": True,
        "description": "Earlier Qwen generation · native vision",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/Qwen/Qwen3.6-27B",
        "pricing_url": "https://router.huggingface.co/v1/models",
    },
    "gemma4-31b": {
        "key": "gemma4-31b",
        "name": "Gemma 4 31B",
        "model_id": "google/gemma-4-31B-it",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "google/gemma-4-31B-it:novita",
        "input_rate": 0.14,
        "output_rate": 0.4,
        "cached_input_rate": 0.14,
        "context_length": 262144,
        "concurrency": 8,
        "generation_options": {},
        "featured": True,
        "description": "Compact vision model · lower token cost",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/google/gemma-4-31B-it",
        "pricing_url": "https://novita.ai/pricing",
    },
    "qwen38-flash-next": {
        "key": "qwen38-flash-next",
        "name": "Qwen3.8 Flash Next",
        "model_id": "Qwen/Qwen3.8-Flash-Next",
        "provider": "featherless-ai",
        "provider_name": "Featherless",
        "route": "Qwen/Qwen3.8-Flash-Next:featherless-ai",
        "input_rate": 0.15,
        "output_rate": 0.5,
        "cached_input_rate": 0.03,
        "context_length": 262144,
        "concurrency": 2,
        "generation_options": {"chat_template_kwargs": {"enable_thinking": False}},
        "featured": False,
        "description": "Original action-phase prompt development model",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "2 simultaneous requests tested on this account",
        "model_url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
        "pricing_url": "https://featherless.ai/models/Qwen/Qwen3.8-Flash-Next",
    },
    "glm53-flash": {
        "key": "glm53-flash",
        "name": "GLM 5.3 Flash",
        "model_id": "zai-org/GLM-5.3-Flash",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "zai-org/GLM-5.3-Flash:novita",
        "input_rate": 0.15,
        "output_rate": 0.5,
        "cached_input_rate": 0.03,
        "context_length": 1048576,
        "concurrency": 8,
        "generation_options": {"thinking": {"type": "disabled"}},
        "featured": False,
        "description": "Low-cost multimodal option",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/zai-org/GLM-5.3-Flash",
        "pricing_url": "https://novita.ai/pricing",
    },
    "kimi-k3": {
        "key": "kimi-k3",
        "name": "Kimi K3",
        "model_id": "moonshotai/Kimi-K3",
        "provider": "baseten",
        "provider_name": "Baseten",
        "route": "moonshotai/Kimi-K3:baseten",
        "input_rate": 3,
        "output_rate": 15,
        "cached_input_rate": 0.3,
        "context_length": 1048576,
        "concurrency": 8,
        "generation_options": {"reasoning_effort": "none"},
        "featured": False,
        "description": "Premium multimodal option · higher token cost",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/moonshotai/Kimi-K3",
        "pricing_url": "https://www.baseten.co/library/kimi-k3/",
    },
    "minimax-m3": {
        "key": "minimax-m3",
        "name": "MiniMax M3",
        "model_id": "MiniMaxAI/MiniMax-M3",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "MiniMaxAI/MiniMax-M3:novita",
        "input_rate": 0.3,
        "output_rate": 1.2,
        "cached_input_rate": 0.06,
        "context_length": 1000000,
        "concurrency": 8,
        "generation_options": {},
        "featured": False,
        "description": "Multimodal option with fast generation",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3",
        "pricing_url": "https://novita.ai/pricing",
        "pricing_tiers": [
            {"min_input_tokens": 524288, "input_rate": 0.6, "output_rate": 2.4, "cached_input_rate": 0.12}
        ],
        "pricing_note": "Base rates shown. Requests with 524,288 or more input tokens cost twice as much.",
    },
    "gemma4-26b": {
        "key": "gemma4-26b",
        "name": "Gemma 4 26B A4B",
        "model_id": "google/gemma-4-26B-A4B-it",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "google/gemma-4-26B-A4B-it:novita",
        "input_rate": 0.13,
        "output_rate": 0.4,
        "cached_input_rate": 0.13,
        "context_length": 262144,
        "concurrency": 8,
        "generation_options": {},
        "featured": False,
        "description": "Efficient mixture-of-experts vision model",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/google/gemma-4-26B-A4B-it",
        "pricing_url": "https://novita.ai/pricing",
    },
    "qwen3-vl-235b": {
        "key": "qwen3-vl-235b",
        "name": "Qwen3 VL 235B",
        "model_id": "Qwen/Qwen3-VL-235B-A22B-Instruct",
        "provider": "novita",
        "provider_name": "Novita",
        "route": "Qwen/Qwen3-VL-235B-A22B-Instruct:novita",
        "input_rate": 0.3,
        "output_rate": 1.5,
        "cached_input_rate": 0.3,
        "context_length": 131072,
        "concurrency": 8,
        "generation_options": {},
        "featured": False,
        "description": "Large dedicated vision-language model",
        "checked_at": "2026-09-25",
        "vision_test": "Two-image JSON smoke test passed",
        "concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
        "model_url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct",
        "pricing_url": "https://novita.ai/pricing",
    },
}

UNAVAILABLE = []

# Robot → human demonstration videos. Both media models run on fal through HF Inference Providers
# and are billed per request to the signed-in account, so the Space reserves a fixed quoted price.
HUMAN_VIDEO = {
    "provider": "fal-ai",
    "provider_name": "fal",
    "edit_model": "Qwen/Qwen-Image-Edit-2509",
    # Lower guidance keeps the scene's colours; fal's default oversaturates.
    "edit_guidance_scale": 1.5,
    "video_model": "MiniMaxAI/MiniMax-H3",
    # MiniMax-H3 is mapped under the image-text-to-video task on fal.
    "video_task": "image-text-to-video",
    "resolution": "480P",
    "duration_s": 5,
    # fal list prices. The edit is billed per output megapixel (fal returns 1024x1024), the video
    # per second of output.
    "edit_usd_per_megapixel": 0.03,
    "edit_megapixels": 1.05,
    "video_usd_per_second": 0.05,
    # Reserved before each request and reconciled to the price above on success.
    "edit_usd": 0.04,
    "video_usd": 0.30,
    # Default simultaneous media requests per job (user-selectable up to 32).
    "concurrency": 8,
    # Estimation only, calibrated on the first Studio video run (3 ReBot episodes, 159.7 s,
    # 33 subtasks): subtasks ≈ 1 per episode + 1 per 5.3 s. The two VLM calls per video used
    # about 1,750 input and 550 output tokens together. A video takes up to ~3 min on fal.
    "subtasks_per_episode": 1,
    "seconds_per_subtask": 5.3,
    "vlm_input_tokens": 1750,
    "vlm_output_tokens": 550,
    "minutes_per_video": 3,
    "checked_at": "2026-09-30",
    "pricing_url": "https://fal.ai/models/minimax/h3/image-to-video",
}
# 0 = one video for every subtask.
HUMAN_VIDEO_SEGMENT_CHOICES = (0, 1, 2, 3, 5, 8)
HUMAN_VIDEO_PARALLEL_CHOICES = (1, 4, 8, 16, 32)


def model_profile(key=DEFAULT_MODEL):
    if key not in MODELS:
        raise ValueError("Choose a supported vision model from the model selector.")
    return deepcopy(MODELS[key])


def human_video_profile():
    return deepcopy(HUMAN_VIDEO)


def human_video_segment_cost(profile=None):
    """Expected fal charge for one video (edit + animation) at list prices."""
    profile = profile or HUMAN_VIDEO
    return (
        profile["edit_usd_per_megapixel"] * profile["edit_megapixels"]
        + profile["video_usd_per_second"] * profile["duration_s"]
    )


def human_video_segment_reservation(profile=None):
    """Conservative amount held from the spending target while one video is in flight."""
    profile = profile or HUMAN_VIDEO
    return profile["edit_usd"] + profile["video_usd"]


def public_models():
    return {
        "default": DEFAULT_MODEL,
        "models": [model_profile(k) for k in MODELS],
        "unavailable": deepcopy(UNAVAILABLE),
        "human_video": {
            **human_video_profile(),
            "segment_choices": list(HUMAN_VIDEO_SEGMENT_CHOICES),
            "parallel_choices": list(HUMAN_VIDEO_PARALLEL_CHOICES),
            "segment_usd": human_video_segment_cost(),
        },
    }

	"""Curated, vision-tested HF routes. Server-owned pricing and request parameters.

	This is a practical selection, not a universal quality ranking. Keep checked_at
	and smoke tests current when changing any route. Never infer vision support from
	model family names or silently substitute a similarly named checkpoint.
	"""

	from copy import deepcopy

	DEFAULT_MODEL = "qwen38-27b"
	MODELS = {
	"qwen38-27b": {
	"key": "qwen38-27b",
	"name": "Qwen3.8 27B",
	"model_id": "Qwen/Qwen3.8-27B",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "Qwen/Qwen3.8-27B:novita",
	"input_rate": 0.42,
	"output_rate": 3,
	"cached_input_rate": 0.085,
	"context_length": 1000000,
	"concurrency": 32,
	"generation_options": {"chat_template_kwargs": {"enable_thinking": False}},
	"featured": True,
	"description": "Current default · tested on 33 robot episodes",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "32 simultaneous vision requests tested",
	"model_url": "https://huggingface.co/Qwen/Qwen3.8-27B",
	"pricing_url": "https://novita.ai/pricing",
	},
	"qwen36-27b": {
	"key": "qwen36-27b",
	"name": "Qwen3.6 27B",
	"model_id": "Qwen/Qwen3.6-27B",
	"provider": "ovhcloud",
	"provider_name": "OVHcloud",
	"route": "Qwen/Qwen3.6-27B:ovhcloud",
	"input_rate": 0.47,
	"output_rate": 3.19,
	"cached_input_rate": 0.47,
	"context_length": 262144,
	"concurrency": 8,
	"generation_options": {"reasoning_effort": "none"},
	"featured": True,
	"description": "Earlier Qwen generation · native vision",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/Qwen/Qwen3.6-27B",
	"pricing_url": "https://router.huggingface.co/v1/models",
	},
	"gemma4-31b": {
	"key": "gemma4-31b",
	"name": "Gemma 4 31B",
	"model_id": "google/gemma-4-31B-it",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "google/gemma-4-31B-it:novita",
	"input_rate": 0.14,
	"output_rate": 0.4,
	"cached_input_rate": 0.14,
	"context_length": 262144,
	"concurrency": 8,
	"generation_options": {},
	"featured": True,
	"description": "Compact vision model · lower token cost",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/google/gemma-4-31B-it",
	"pricing_url": "https://novita.ai/pricing",
	},
	"qwen38-flash-next": {
	"key": "qwen38-flash-next",
	"name": "Qwen3.8 Flash Next",
	"model_id": "Qwen/Qwen3.8-Flash-Next",
	"provider": "featherless-ai",
	"provider_name": "Featherless",
	"route": "Qwen/Qwen3.8-Flash-Next:featherless-ai",
	"input_rate": 0.15,
	"output_rate": 0.5,
	"cached_input_rate": 0.03,
	"context_length": 262144,
	"concurrency": 2,
	"generation_options": {"chat_template_kwargs": {"enable_thinking": False}},
	"featured": False,
	"description": "Original action-phase prompt development model",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "2 simultaneous requests tested on this account",
	"model_url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
	"pricing_url": "https://featherless.ai/models/Qwen/Qwen3.8-Flash-Next",
	},
	"glm53-flash": {
	"key": "glm53-flash",
	"name": "GLM 5.3 Flash",
	"model_id": "zai-org/GLM-5.3-Flash",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "zai-org/GLM-5.3-Flash:novita",
	"input_rate": 0.15,
	"output_rate": 0.5,
	"cached_input_rate": 0.03,
	"context_length": 1048576,
	"concurrency": 8,
	"generation_options": {"thinking": {"type": "disabled"}},
	"featured": False,
	"description": "Low-cost multimodal option",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/zai-org/GLM-5.3-Flash",
	"pricing_url": "https://novita.ai/pricing",
	},
	"kimi-k3": {
	"key": "kimi-k3",
	"name": "Kimi K3",
	"model_id": "moonshotai/Kimi-K3",
	"provider": "baseten",
	"provider_name": "Baseten",
	"route": "moonshotai/Kimi-K3:baseten",
	"input_rate": 3,
	"output_rate": 15,
	"cached_input_rate": 0.3,
	"context_length": 1048576,
	"concurrency": 8,
	"generation_options": {"reasoning_effort": "none"},
	"featured": False,
	"description": "Premium multimodal option · higher token cost",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/moonshotai/Kimi-K3",
	"pricing_url": "https://www.baseten.co/library/kimi-k3/",
	},
	"minimax-m3": {
	"key": "minimax-m3",
	"name": "MiniMax M3",
	"model_id": "MiniMaxAI/MiniMax-M3",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "MiniMaxAI/MiniMax-M3:novita",
	"input_rate": 0.3,
	"output_rate": 1.2,
	"cached_input_rate": 0.06,
	"context_length": 1000000,
	"concurrency": 8,
	"generation_options": {},
	"featured": False,
	"description": "Multimodal option with fast generation",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/MiniMaxAI/MiniMax-M3",
	"pricing_url": "https://novita.ai/pricing",
	"pricing_tiers": [
	{"min_input_tokens": 524288, "input_rate": 0.6, "output_rate": 2.4, "cached_input_rate": 0.12}
	],
	"pricing_note": "Base rates shown. Requests with 524,288 or more input tokens cost twice as much.",
	},
	"gemma4-26b": {
	"key": "gemma4-26b",
	"name": "Gemma 4 26B A4B",
	"model_id": "google/gemma-4-26B-A4B-it",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "google/gemma-4-26B-A4B-it:novita",
	"input_rate": 0.13,
	"output_rate": 0.4,
	"cached_input_rate": 0.13,
	"context_length": 262144,
	"concurrency": 8,
	"generation_options": {},
	"featured": False,
	"description": "Efficient mixture-of-experts vision model",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/google/gemma-4-26B-A4B-it",
	"pricing_url": "https://novita.ai/pricing",
	},
	"qwen3-vl-235b": {
	"key": "qwen3-vl-235b",
	"name": "Qwen3 VL 235B",
	"model_id": "Qwen/Qwen3-VL-235B-A22B-Instruct",
	"provider": "novita",
	"provider_name": "Novita",
	"route": "Qwen/Qwen3-VL-235B-A22B-Instruct:novita",
	"input_rate": 0.3,
	"output_rate": 1.5,
	"cached_input_rate": 0.3,
	"context_length": 131072,
	"concurrency": 8,
	"generation_options": {},
	"featured": False,
	"description": "Large dedicated vision-language model",
	"checked_at": "2026-09-25",
	"vision_test": "Two-image JSON smoke test passed",
	"concurrency_note": "Conservative setting; sustained concurrency not benchmarked",
	"model_url": "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct",
	"pricing_url": "https://novita.ai/pricing",
	},
	}

	UNAVAILABLE = []

	# Robot → human demonstration videos. Both media models run on fal through HF Inference Providers
	# and are billed per request to the signed-in account, so the Space reserves a fixed quoted price.
	HUMAN_VIDEO = {
	"provider": "fal-ai",
	"provider_name": "fal",
	"edit_model": "Qwen/Qwen-Image-Edit-2509",
	# Lower guidance keeps the scene's colours; fal's default oversaturates.
	"edit_guidance_scale": 1.5,
	"video_model": "MiniMaxAI/MiniMax-H3",
	# MiniMax-H3 is mapped under the image-text-to-video task on fal.
	"video_task": "image-text-to-video",
	"resolution": "480P",
	"duration_s": 5,
	# fal list prices. The edit is billed per output megapixel (fal returns 1024x1024), the video
	# per second of output.
	"edit_usd_per_megapixel": 0.03,
	"edit_megapixels": 1.05,
	"video_usd_per_second": 0.05,
	# Reserved before each request and reconciled to the price above on success.
	"edit_usd": 0.04,
	"video_usd": 0.30,
	# Default simultaneous media requests per job (user-selectable up to 32).
	"concurrency": 8,
	# Estimation only, calibrated on the first Studio video run (3 ReBot episodes, 159.7 s,
	# 33 subtasks): subtasks ≈ 1 per episode + 1 per 5.3 s. The two VLM calls per video used
	# about 1,750 input and 550 output tokens together. A video takes up to ~3 min on fal.
	"subtasks_per_episode": 1,
	"seconds_per_subtask": 5.3,
	"vlm_input_tokens": 1750,
	"vlm_output_tokens": 550,
	"minutes_per_video": 3,
	"checked_at": "2026-09-30",
	"pricing_url": "https://fal.ai/models/minimax/h3/image-to-video",
	}
	# 0 = one video for every subtask.
	HUMAN_VIDEO_SEGMENT_CHOICES = (0, 1, 2, 3, 5, 8)
	HUMAN_VIDEO_PARALLEL_CHOICES = (1, 4, 8, 16, 32)


	def model_profile(key=DEFAULT_MODEL):
	if key not in MODELS:
	raise ValueError("Choose a supported vision model from the model selector.")
	return deepcopy(MODELS[key])


	def human_video_profile():
	return deepcopy(HUMAN_VIDEO)


	def human_video_segment_cost(profile=None):
	"""Expected fal charge for one video (edit + animation) at list prices."""
	profile = profile or HUMAN_VIDEO
	return (
	profile["edit_usd_per_megapixel"] * profile["edit_megapixels"]
	+ profile["video_usd_per_second"] * profile["duration_s"]
	)


	def human_video_segment_reservation(profile=None):
	"""Conservative amount held from the spending target while one video is in flight."""
	profile = profile or HUMAN_VIDEO
	return profile["edit_usd"] + profile["video_usd"]


	def public_models():
	return {
	"default": DEFAULT_MODEL,
	"models": [model_profile(k) for k in MODELS],
	"unavailable": deepcopy(UNAVAILABLE),
	"human_video": {
	**human_video_profile(),
	"segment_choices": list(HUMAN_VIDEO_SEGMENT_CHOICES),
	"parallel_choices": list(HUMAN_VIDEO_PARALLEL_CHOICES),
	"segment_usd": human_video_segment_cost(),
	},
	}