Chronos2 forecasting model

[topepo]

This is a pure R torch version of Chronos version 2. It is a pre-trained model; weights are downloaded and cached on first use.

A GPU is not required, but extremely helpful at prediction time. Unlike most other models, this implementation appears to achieve good speedups on Apple GPUs (aka MPS devices).

[github-actions]

Remaining comments which cannot be posted as a review comment to avoid GitHub Rate Limit

air

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 788 to 794 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 829 to 835 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 863 to 869 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 898 to 904 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 931 to 937 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Line 951 in 258f57a

ctx_list <- list(torch::torch_tensor(rnorm(16), dtype = torch::torch_float32()))

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 965 to 971 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 1000 to 1006 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 1022 to 1023 in 258f57a

	tiny_model, tiny_config, torch::torch_device("cpu"),
	inputs, prediction_length = 4L, num_output_patches = 1L

---

[air] _{reported by reviewdog 🐶}

brulee/tests/testthat/test-chronos2-predict.R

Lines 1035 to 1041 in 258f57a

	d_model = 32L, d_ff = 64L, d_kv = 16L, num_heads = 2L,
	num_layers = 1L, dropout_rate = 0.0, layer_norm_epsilon = 1e-6,
	rope_theta = 10000, vocab_size = 2L, pad_token_id = 0L,
	reg_token_id = 1L, context_length = 64L, input_patch_size = 4L,
	input_patch_stride = 4L, output_patch_size = 4L,
	max_output_patches = 4L, quantiles = c(0.1, 0.5, 0.9),
	use_arcsinh = FALSE, use_reg_token = TRUE,

[topepo]

Maybe we change this along the way as more versions are populated. For now, it's keyed to this particular version.

[topepo]

R's internal file.download() had a lot of issues with downloading this file, so we went with curl.

[topepo]

This seems like a good place to put the weights. I'm not aware of there is a canonical location for cached objects related to R.