From ea20f2be0fdd6bd5ac1a2e603f63614a4a983ec9 Mon Sep 17 00:00:00 2001 From: Jagadish Krishnamoorthy Date: Wed, 6 Dec 2023 00:30:37 -0800 Subject: [PATCH 1/7] Update README.md Fixed typo --- transformer_lm/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/transformer_lm/README.md b/transformer_lm/README.md index 16997197..d14302a5 100644 --- a/transformer_lm/README.md +++ b/transformer_lm/README.md @@ -11,4 +11,4 @@ python main.py --gpu By default the dataset is the [PTB corpus](https://paperswithcode.com/dataset/penn-treebank). Choose a different dataset with the `--dataset` option. -To run the PyTorch, Jax or TensorFlowexamples install the respective framework. +To run the PyTorch, Jax or TensorFlow examples install the respective framework. From a7c5beb1a748da55725cf9778122b3087bc0028f Mon Sep 17 00:00:00 2001 From: Vaibhav Srivastav Date: Wed, 6 Dec 2023 19:13:14 +0100 Subject: [PATCH 2/7] Add url to mlx checkpoints on HF. --- llama/README.md | 3 +++ 1 file changed, 3 insertions(+) diff --git a/llama/README.md b/llama/README.md index 1314ca86..63aba7d9 100644 --- a/llama/README.md +++ b/llama/README.md @@ -17,6 +17,9 @@ weights you will need to [request access](https://docs.google.com/forms/d/e/1FAIpQLSfqNECQnMkycAp2jP4Z9TFX0cGR4uf7b_fBxjY_OjhJILlKGA/viewform) from Meta. + +Alternatively, you can also download a select converted checkpoints from the [mlx-llama](https://huggingface.co/mlx-llama) community organisation on Hugging Face and skip the conversion step. + Convert the weights with: ``` From 22e067648e19af015c278272286a0bd3d16c069c Mon Sep 17 00:00:00 2001 From: adhishthite Date: Thu, 7 Dec 2023 00:07:42 +0530 Subject: [PATCH 3/7] Benchmark all models if user allows. --- .gitignore | 2 + whisper/benchmark.py | 99 +++++++++++++++++++++++++------------------- 2 files changed, 58 insertions(+), 43 deletions(-) diff --git a/.gitignore b/.gitignore index b6e47617..51288c78 100644 --- a/.gitignore +++ b/.gitignore @@ -127,3 +127,5 @@ dmypy.json # Pyre type checker .pyre/ +.idea/ +.vscode/ diff --git a/whisper/benchmark.py b/whisper/benchmark.py index 880453a0..9df6b500 100644 --- a/whisper/benchmark.py +++ b/whisper/benchmark.py @@ -1,5 +1,6 @@ # Copyright © 2023 Apple Inc. +import sys import time import mlx.core as mx @@ -48,46 +49,58 @@ def everything(): if __name__ == "__main__": - feat_time = timer(feats) - print(f"Feature time {feat_time:.3f}") - mels = feats()[None] - tokens = mx.array( - [ - 50364, - 1396, - 264, - 665, - 5133, - 23109, - 25462, - 264, - 6582, - 293, - 750, - 632, - 42841, - 292, - 370, - 938, - 294, - 4054, - 293, - 12653, - 356, - 50620, - 50620, - 23563, - 322, - 3312, - 13, - 50680, - ], - mx.int32, - )[None] - model = load_models.load_model("tiny") - model_forward_time = timer(model_forward, model, mels, tokens) - print(f"Model forward time {model_forward_time:.3f}") - decode_time = timer(decode, model, mels) - print(f"Decode time {decode_time:.3f}") - everything_time = timer(everything) - print(f"Everything time {everything_time:.3f}") + + # get command line arguments without 3rd party libraries + # the command line argument to benchmark all models is "all" + models = ["tiny"] + if len(sys.argv) > 1: + if sys.argv[1] == "--all": + models = ["tiny", "small", "medium", "large"] + + for model_name in models: + feat_time = timer(feats) + + print(f"\nModel: {model_name.upper()}") + print(f"\nFeature time {feat_time:.3f}") + mels = feats()[None] + tokens = mx.array( + [ + 50364, + 1396, + 264, + 665, + 5133, + 23109, + 25462, + 264, + 6582, + 293, + 750, + 632, + 42841, + 292, + 370, + 938, + 294, + 4054, + 293, + 12653, + 356, + 50620, + 50620, + 23563, + 322, + 3312, + 13, + 50680, + ], + mx.int32, + )[None] + model = load_models.load_model(f"{model_name}") + model_forward_time = timer(model_forward, model, mels, tokens) + print(f"Model forward time {model_forward_time:.3f}") + decode_time = timer(decode, model, mels) + print(f"Decode time {decode_time:.3f}") + everything_time = timer(everything) + print(f"Everything time {everything_time:.3f}") + print(f"\n{'-----' * 10}\n") From 3b7bae0b93182be9cc28ceddca7294519846a362 Mon Sep 17 00:00:00 2001 From: Hitarth Sharma Date: Thu, 7 Dec 2023 08:21:21 +1300 Subject: [PATCH 4/7] Add missing numpy dependency --- mistral/requirements.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/mistral/requirements.txt b/mistral/requirements.txt index 7111f1d4..d775b88f 100644 --- a/mistral/requirements.txt +++ b/mistral/requirements.txt @@ -1,3 +1,4 @@ mlx sentencepiece torch +numpy From c693bb75b93e4144bf180f789aa6f0ff45e6524e Mon Sep 17 00:00:00 2001 From: Leon Ericsson Date: Wed, 6 Dec 2023 17:42:23 +0100 Subject: [PATCH 5/7] bug fix in transformer_lm example --- transformer_lm/main.py | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-) diff --git a/transformer_lm/main.py b/transformer_lm/main.py index f94003fd..b0d2966b 100644 --- a/transformer_lm/main.py +++ b/transformer_lm/main.py @@ -81,13 +81,13 @@ def main(args): optimizer = optim.SGD(learning_rate=args.learning_rate) loss_and_grad_fn = nn.value_and_grad(model, model.loss) - def eval_fn(params, dataset): + def eval_fn(model, dataset): inputs, targets = map(mx.array, to_samples(context_size, dataset)) loss = 0 for s in range(0, targets.shape[0], batch_size): bx, by = inputs[s : s + batch_size], targets[s : s + batch_size] bx, by = map(mx.array, (bx, by)) - losses = self.loss(bx, by, reduce=False) + losses = model.loss(bx, by, reduce=False) loss += mx.sum(losses).item() return loss / len(targets) @@ -110,9 +110,8 @@ def main(args): ) losses = [] tic = time.perf_counter() - if (it + 1) % steps_per_eval == 0: - val_loss = eval_fn(params, valid) + val_loss = eval_fn(model, valid) toc = time.perf_counter() print( f"Iter {it + 1}: " @@ -123,7 +122,7 @@ def main(args): tic = time.perf_counter() if args.eval_test: - test_loss = eval_fn(params, test) + test_loss = eval_fn(model, test) test_ppl = math.exp(test_loss) print(f"Test loss {test_loss:.3f}, Test ppl {test_ppl:.3f}.") From 5cc000082e241f9c7e9124a4c93e6f9209daf97d Mon Sep 17 00:00:00 2001 From: Angelos Katharopoulos Date: Wed, 6 Dec 2023 15:22:46 -0800 Subject: [PATCH 6/7] Clarification in the stable diffusion readme --- stable_diffusion/README.md | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/stable_diffusion/README.md b/stable_diffusion/README.md index 170ce6b9..edd3a0ed 100644 --- a/stable_diffusion/README.md +++ b/stable_diffusion/README.md @@ -65,8 +65,9 @@ Performance ----------- The following table compares the performance of the UNet in stable diffusion. -We report throughput in images per second for the provided `txt2image.py` -script and the `diffusers` library using the MPS PyTorch backend. +We report throughput in images per second **processed by the UNet** for the +provided `txt2image.py` script and the `diffusers` library using the MPS +PyTorch backend. At the time of writing this comparison convolutions are still some of the least optimized operations in MLX. Despite that, MLX still achieves **~40% higher @@ -93,3 +94,7 @@ The above experiments were made on an M2 Ultra with PyTorch version 2.1, diffusers version 0.21.4 and transformers version 4.33.3. For the generation we used classifier free guidance which means that the above batch sizes result double the images processed by the UNet. + +Note that the above table means that it takes about 90 seconds to fully +generate 16 images with MLX and 50 diffusion steps with classifier free +guidance and about 120 for PyTorch. From 0aea0f3674c428a04b980b21625187f49516a1f5 Mon Sep 17 00:00:00 2001 From: David Chavez Date: Wed, 6 Dec 2023 12:08:43 +0100 Subject: [PATCH 7/7] fix(sd): Add mlx as a dependency --- stable_diffusion/requirements.txt | 1 + 1 file changed, 1 insertion(+) diff --git a/stable_diffusion/requirements.txt b/stable_diffusion/requirements.txt index 7d3b8f5d..c2fa7225 100644 --- a/stable_diffusion/requirements.txt +++ b/stable_diffusion/requirements.txt @@ -1,3 +1,4 @@ +mlx safetensors huggingface-hub regex