ChronosBoltPipeline.predict_quantiles takes `inputs` positional, not
`context` keyword. Use positional with TypeError fallback for older
chronos versions.
FP16 caused inf overflow on Korean stock prices (e.g. 280,000원 >
FP16 max 65,504). Force FP32 for prices to avoid this. Chronos model
itself handles internal scaling.
Verified end-to-end: 60-day daily fetch → Chronos predict → quantile
output. Example 005930: median=-0.59%, q10=-8.9%, q90=+6.4%, conf=0.0
(low conf is mathematically correct when median is near zero relative
to distribution width).
45/45 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ChronosBoltPipeline.predict() does not accept `context` kwarg; it
uses positional-only and is deterministic (no num_samples). Switch
to predict_quantiles(context, prediction_length, quantile_levels)
which returns (quantiles_tensor, mean_tensor).
Implementation: if hasattr(pipeline, "predict_quantiles") → modern
quantile branch. Else fall back to legacy sample-based predict (T5).
Tests: switch to predict_quantiles mock returning (quantiles, None)
with shape [1, 1, 3] for q10/q50/q90 directly.
45/45 tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ChronosPipeline (legacy T5) does not support amazon/chronos-2 or
chronos-bolt-* (input_patch_size). Switch to BaseChronosPipeline
which auto-detects variant and returns the appropriate sub-pipeline
(ChronosBoltPipeline / Chronos2Pipeline / ChronosPipeline).
Also handle the dtype kwarg deprecation: try newer `dtype=` first,
fall back to `torch_dtype=` for older versions.
Test mock_pipeline fixture updated to patch BaseChronosPipeline.
45/45 tests pass. Verified amazon/chronos-bolt-base loads on CUDA.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
