mudler/LocalAI · warning · RuntimeError

request was cancelled

Error message

request was cancelled

What it means

RuntimeError raised inside the avatar segment loop when context.is_active() reports the gRPC context is no longer active — the client cancelled the call, its deadline expired, or the connection dropped. Because avatar generation is sequential over segments (each ~93 frames with 13 conditioning frames carried over), the loop checks cancellation between segments and aborts instead of burning GPU on a dead request. Unlike the ValueError checks, this is a server-side cancellation signal, not a client input error.

Source

Thrown at backend/python/longcat-video/backend.py:729

                **common,
            )
        else:
            width, height = validate_dimensions(request.width, request.height)
            output, latent = self.pipeline.generate_at2v(
                height=height,
                width=width,
                **common,
            )

        video = self._frames_to_pil(output[0])
        width, height = video[0].size
        current_video = video
        reference_latent = latent[:, :, :1].clone()
        all_frames = list(video)

        for segment in range(1, segments):
            if hasattr(context, "is_active") and not context.is_active():
                raise RuntimeError("request was cancelled")
            print(
                f"Generating avatar segment {segment + 1}/{segments}", file=sys.stderr
            )
            audio_start += segment_frames - conditioning_frames
            output, latent = self.pipeline.generate_avc(
                video=current_video,
                video_latent=latent,
                prompt=request.prompt,
                negative_prompt=negative_prompt,
                height=height,
                width=width,
                num_frames=segment_frames,
                num_cond_frames=conditioning_frames,
                num_inference_steps=steps,
                text_guidance_scale=text_guidance,
                audio_guidance_scale=audio_guidance,
                generator=generator,
                output_type="both",

View on GitHub (pinned to 44413a9d06)

Solutions

1. Raise or remove the client-side deadline so it exceeds worst-case total generation time (segments x per-segment time)
2. Only cancel when truly intended; if it fires spuriously, keep the gRPC channel alive (heartbeats/keepalive) during long jobs
3. Treat this error as benign cleanup on the server; on the client map UNAVAILABLE/CANCELLED to a user-facing 'generation stopped' state

Example fix

# before
response = stub.Video gen(timeout=30))  # too short for multi-segment

# after
response = stub.GenerateVideo(req, timeout=1800)  # deadline sized to full job

# No pre-validation possible (server-observed cancellation), but size the deadline:
# estimated segments = ceil((audio_seconds * 25 - 93) / 80) + 1
# deadline > segments * worst_case_seconds_per_segment + slack
import math
segments = max(1, math.ceil((audio_seconds * 25 - 93) / 80) + 1)
timeout = int(segments * 120) + 60

try:
    resp = stub.GenerateVideo(req, timeout=timeout)
except grpc.RpcError as e:
    if e.code() in (grpc.StatusCode.CANCELLED, grpc.StatusCode.DEADLINE_EXCEEDED):
        log.info("generation cancelled/deadline — retry with longer timeout if wanted")
        # safe to retry: generation is stateless per request
    else:
        raise

When it happens

Trigger: Client calls context.cancel() or closes the stream mid-generation; per-call deadline (timeout) shorter than total generation time for all segments; network drop between client and backend.

Common situations: UI 'stop' button cancelling generation; deadline set for short requests reused on long multi-segment avatar jobs; load balancer idle-cutting long-running streams.

Related errors

• Model not loaded
• start_image is not a readable staged file
• num_frames must not be negative
• audio input requires a LongCat-Video-Avatar-1.5 model
• audio is required for LongCat-Video-Avatar-1.5

AI-assisted analysis of mudler/LocalAI@44413a9d06 (2026-08-15). Data as JSON: /api/errors/fbf8a6f2e85ddfc3. Report an issue: GitHub.