liger-kernel 0.6.2__tar.gz → 0.6.4__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/amd-ci.yml +2 -4
- liger_kernel-0.6.4/.github/workflows/docs.yml +64 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/intel-ci.yml +0 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/nvi-ci.yml +0 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.gitignore +2 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/Makefile +9 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/PKG-INFO +16 -10
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/README.md +12 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/data/all_benchmark_data.csv +304 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_cross_entropy.py +4 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_fused_linear_cross_entropy.py +25 -19
- liger_kernel-0.6.4/benchmark/scripts/benchmark_grpo_loss.py +234 -0
- liger_kernel-0.6.4/benchmark/scripts/benchmark_poly_norm.py +197 -0
- liger_kernel-0.6.4/benchmark/scripts/benchmark_tiled_mlp.py +397 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/dev/modal/tests.py +1 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/dev/modal/tests_bwd.py +1 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/Examples.md +1 -1
- liger_kernel-0.6.4/docs/High-Level-APIs.md +93 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/acknowledgement.md +0 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/alignment/run_orpo.py +1 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/training.py +2 -2
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/training_multimodal.py +1 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/train.py +1 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/mkdocs.yml +27 -9
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/pyproject.toml +29 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/setup.py +14 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/cosine_similarity_loss.py +13 -4
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/fused_linear_distillation.py +13 -2
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/fused_linear_ppo.py +25 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/grpo_loss.py +46 -9
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/jsd_loss.py +23 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/cross_entropy.py +118 -62
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/fused_linear_cross_entropy.py +97 -13
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/grpo_loss.py +3 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/layer_norm.py +86 -69
- liger_kernel-0.6.4/src/liger_kernel/ops/poly_norm.py +386 -0
- liger_kernel-0.6.4/src/liger_kernel/ops/tiled_mlp.py +136 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/__init__.py +36 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/cross_entropy.py +8 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/functional.py +31 -6
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/fused_linear_cross_entropy.py +13 -4
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/grpo_loss.py +56 -1
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/falcon_h1.py +122 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/gemma.py +19 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/gemma2.py +22 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/gemma3.py +52 -14
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/glm4.py +18 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/glm4v.py +19 -6
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/glm4v_moe.py +172 -0
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/hunyuan_v1.py +134 -0
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/internvl.py +157 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/llama.py +16 -6
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/llama4.py +18 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/llava.py +18 -6
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/loss_utils.py +32 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/mistral.py +17 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/mixtral.py +24 -9
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/mllama.py +14 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/olmo2.py +18 -5
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/olmo3.py +142 -0
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/output_classes.py +147 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/paligemma.py +41 -5
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/phi3.py +16 -8
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/qwen2.py +18 -4
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/qwen2_5_vl.py +163 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/qwen2_vl.py +24 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/qwen3.py +22 -6
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/qwen3_moe.py +27 -7
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/qwen3_next.py +146 -0
- liger_kernel-0.6.2/src/liger_kernel/transformers/model/qwen2_5_vl.py → liger_kernel-0.6.4/src/liger_kernel/transformers/model/qwen3_vl.py +17 -17
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/qwen3_vl_moe.py +126 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/smollm3.py +17 -7
- liger_kernel-0.6.4/src/liger_kernel/transformers/model/smolvlm.py +158 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/monkey_patch.py +830 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/multi_token_attention.py +1 -1
- liger_kernel-0.6.4/src/liger_kernel/transformers/poly_norm.py +42 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/rms_norm.py +7 -0
- liger_kernel-0.6.4/src/liger_kernel/transformers/rope.py +63 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/swiglu.py +17 -0
- liger_kernel-0.6.4/src/liger_kernel/transformers/tiled_mlp.py +133 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel.egg-info/PKG-INFO +16 -10
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel.egg-info/SOURCES.txt +22 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel.egg-info/requires.txt +3 -4
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_grpo_loss.py +234 -26
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_jsd_loss.py +5 -2
- liger_kernel-0.6.4/test/conftest.py +11 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/bf16/test_mini_models.py +660 -2
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/bf16/test_mini_models_multimodal.py +420 -3
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/bf16/test_mini_models_with_logits.py +656 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/fp32/test_mini_models.py +631 -1
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/fp32/test_mini_models_multimodal.py +559 -7
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/fp32/test_mini_models_with_logits.py +651 -0
- liger_kernel-0.6.4/test/resources/fake_configs/HuggingFaceTB/SmolVLM2-256M-Video-Instruct/tokenizer_config.json +1192 -0
- liger_kernel-0.6.4/test/resources/fake_configs/OpenGVLab/InternVL3-1B-hf/tokenizer_config.json +307 -0
- liger_kernel-0.6.4/test/resources/fake_configs/Qwen/Qwen3-VL-4B-Instruct/tokenizer_config.json +63 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_cross_entropy.py +126 -6
- liger_kernel-0.6.4/test/transformers/test_fused_linear_cross_entropy.py +955 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_layer_norm.py +1 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_monkey_patch.py +1395 -218
- liger_kernel-0.6.4/test/transformers/test_poly_norm.py +281 -0
- liger_kernel-0.6.4/test/transformers/test_tiled_mlp.py +216 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/utils.py +157 -7
- liger_kernel-0.6.2/.github/workflows/docs.yml +0 -30
- liger_kernel-0.6.2/docs/High-Level-APIs.md +0 -30
- liger_kernel-0.6.2/src/liger_kernel/transformers/rope.py +0 -20
- liger_kernel-0.6.2/test/conftest.py +0 -8
- liger_kernel-0.6.2/test/transformers/test_fused_linear_cross_entropy.py +0 -354
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/ISSUE_TEMPLATE/bug_report.yaml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/ISSUE_TEMPLATE/feature_request.yaml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/pull_request_template.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/benchmark.yml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/publish-nightly.yml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/.github/workflows/publish-release.yml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/LICENSE +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/NOTICE +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/README.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/benchmarks_visualizer.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_cpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_distill_cosine_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_distill_jsd_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_dpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_dyt.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_embedding.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_fused_add_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_fused_linear_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_fused_neighborhood_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_geglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_group_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_kl_div.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_kto_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_layer_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_llama4_rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_multi_token_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_orpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_qwen2vl_mrope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_simpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_softmax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_sparse_multi_token_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_sparsemax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_swiglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/benchmark_tvd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/benchmark/scripts/utils.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/dev/fmt-requirements.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/dev/modal/benchmarks.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/Getting-Started.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/Low-Level-APIs.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/contributing.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/banner.GIF +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/compose.gif +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/e2e-memory.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/e2e-tps.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/logo-banner.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/patch.gif +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/images/post-training.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/index.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/docs/license.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/alignment/accelerate_config.yaml +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/README.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/callback.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/config/fsdp_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/gemma_7b_mem.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/gemma_7b_tp.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/llama_mem_alloc.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/llama_tps.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/qwen_mem_alloc.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/img/qwen_tps.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/launch_on_modal.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/requirements.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/run_benchmarks.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/run_gemma.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/run_llama.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/run_qwen.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/huggingface/run_qwen2_vl.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/lightning/README.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/lightning/requirements.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/lightning/training.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/README.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/callback.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Memory_Stage1_num_head_3.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Memory_Stage1_num_head_5.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Memory_Stage2_num_head_3.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Memory_Stage2_num_head_5.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Throughput_Stage1_num_head_3.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Throughput_Stage1_num_head_5.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Throughput_Stage2_num_head_3.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/docs/images/Throughput_Stage2_num_head_5.png +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/fsdp/acc-fsdp.conf +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/medusa_util.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/requirements.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/examples/medusa/scripts/llama3_8b_medusa.sh +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/licenses/LICENSE-Apache-2.0 +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/licenses/LICENSE-MIT-AutoAWQ +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/licenses/LICENSE-MIT-Efficient-Cross-Entropy +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/licenses/LICENSE-MIT-llmc +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/licenses/LICENSE-MIT-triton +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/setup.cfg +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/README.md +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/cpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/dpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/functional.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/fused_linear_preference.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/fused_linear_unpaired_preference.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/kto_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/orpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/chunked_loss/simpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/env_report.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/dyt.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/experimental/embedding.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/experimental/mm_int8int2.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/fused_add_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/fused_linear_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/fused_neighborhood_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/geglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/group_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/kl_div.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/llama4_rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/multi_token_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/qwen2vl_mrope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/softmax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/sparsemax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/swiglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/tvd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/ops/utils.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/auto_model.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/dyt.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/experimental/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/experimental/embedding.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/fsdp.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/fused_add_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/fused_linear_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/fused_neighborhood_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/geglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/group_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/kl_div.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/layer_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/llama4_rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/model/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/qwen2vl_mrope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/softmax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/sparsemax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/trainer/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/trainer/orpo_trainer.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/trainer_integration.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/transformers/tvd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/triton/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/triton/monkey_patch.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel/utils.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel.egg-info/dependency_links.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/src/liger_kernel.egg-info/top_level.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_cosine_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_cpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_dpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_kto_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_orpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/chunked_loss/test_simpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/bf16/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/convergence/fp32/__init__.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Google/Gemma3/gemma-3-4b-it/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Google/Paligemma/paligemma-3b-pt-224/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Llava/llava-1.5-7b-hf/preprocessor_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Llava/llava-1.5-7b-hf/processor_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Llava/llava-1.5-7b-hf/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Qwen/Qwen2-VL-7B-Instruct/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/Qwen/Qwen2.5-VL-7B-Instruct/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/meta-llama/Llama-3.2-11B-Vision-Instruct/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/fake_configs/meta-llama/Llama-4-Scout-17B-16E-Instruct/tokenizer_config.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/scripts/generate_tokenized_dataset.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/tiny_shakespeare.txt +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/tiny_shakespeare_tokenized/data-00000-of-00001.arrow +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/tiny_shakespeare_tokenized/dataset_info.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/resources/tiny_shakespeare_tokenized/state.json +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_auto_model.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_dyt.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_embedding.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_flex_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_fused_add_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_fused_linear_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_fused_neighborhood_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_geglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_group_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_grpo_loss.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_jsd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_kl_div.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_mm_int8int2.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_multi_token_attention.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_qwen2vl_mrope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_rms_norm.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_rope.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_softmax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_sparsemax.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_swiglu.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_trainer_integration.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_transformers.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/transformers/test_tvd.py +0 -0
- {liger_kernel-0.6.2 → liger_kernel-0.6.4}/test/triton/test_triton_monkey_patch.py +0 -0
|
@@ -13,9 +13,6 @@ on:
|
|
|
13
13
|
paths:
|
|
14
14
|
- "src/**"
|
|
15
15
|
- "test/**"
|
|
16
|
-
schedule:
|
|
17
|
-
# Runs at 00:00 UTC daily
|
|
18
|
-
- cron: '0 0 * * *'
|
|
19
16
|
workflow_dispatch: # Enables manual trigger
|
|
20
17
|
|
|
21
18
|
concurrency:
|
|
@@ -64,7 +61,8 @@ jobs:
|
|
|
64
61
|
run: |
|
|
65
62
|
rocm-smi
|
|
66
63
|
python -m pip install --upgrade pip
|
|
67
|
-
pip install -e .[dev]
|
|
64
|
+
pip install -e .[dev]
|
|
65
|
+
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm_version }}/
|
|
68
66
|
|
|
69
67
|
- name: List Python Environments
|
|
70
68
|
run: python -m pip list
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
name: Publish documentation
|
|
2
|
+
on:
|
|
3
|
+
push:
|
|
4
|
+
branches:
|
|
5
|
+
- main
|
|
6
|
+
paths:
|
|
7
|
+
- 'docs/**'
|
|
8
|
+
- 'mkdocs.yml'
|
|
9
|
+
|
|
10
|
+
permissions:
|
|
11
|
+
contents: write
|
|
12
|
+
jobs:
|
|
13
|
+
deploy:
|
|
14
|
+
runs-on: ubuntu-latest
|
|
15
|
+
steps:
|
|
16
|
+
- uses: actions/checkout@v4
|
|
17
|
+
- name: Configure Git Credentials
|
|
18
|
+
run: |
|
|
19
|
+
git config user.name github-actions[bot]
|
|
20
|
+
git config user.email 41898282+github-actions[bot]@users.noreply.github.com
|
|
21
|
+
- uses: actions/setup-python@v5
|
|
22
|
+
with:
|
|
23
|
+
python-version: 3.x
|
|
24
|
+
- run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV
|
|
25
|
+
- uses: actions/cache@v4
|
|
26
|
+
with:
|
|
27
|
+
key: mkdocs-material-${{ env.cache_id }}
|
|
28
|
+
path: .cache
|
|
29
|
+
restore-keys: |
|
|
30
|
+
mkdocs-material-
|
|
31
|
+
- run: pip install mkdocs-material mkdocstrings[python]
|
|
32
|
+
# ====== Backup the benchmarks from gh-pages ======
|
|
33
|
+
# This is necessary because the benchmarks are not included in the documentation build process.
|
|
34
|
+
# So we need to backup the benchmarks from gh-pages and restore them after the documentation is built.
|
|
35
|
+
- name: Backup benchmarks from gh-pages
|
|
36
|
+
run: |
|
|
37
|
+
git fetch origin gh-pages
|
|
38
|
+
# create worktree bound to local gh-pages, tracking origin/gh-pages
|
|
39
|
+
git branch -f gh-pages origin/gh-pages || true
|
|
40
|
+
mkdir -p ghp && git worktree add ghp gh-pages || true
|
|
41
|
+
if [ -d ghp/benchmarks ]; then
|
|
42
|
+
tar -C ghp -czf /tmp/benchmarks.tgz benchmarks
|
|
43
|
+
fi
|
|
44
|
+
# IMPORTANT: remove worktree so gh-pages isn't checked out anywhere
|
|
45
|
+
git worktree remove ghp --force || true
|
|
46
|
+
echo "Backed up benchmarks from gh-pages"
|
|
47
|
+
# ====== Deploy the documentation ======
|
|
48
|
+
- name: Deploy documentation
|
|
49
|
+
run: mkdocs gh-deploy --force
|
|
50
|
+
# ====== Restore the benchmarks onto gh-pages ======
|
|
51
|
+
# This is necessary because the benchmarks are not included in the documentation build process.
|
|
52
|
+
# So we need to restore the benchmarks onto gh-pages after the documentation is built.
|
|
53
|
+
- name: Restore benchmarks onto gh-pages
|
|
54
|
+
run: |
|
|
55
|
+
# Refresh remote tracking and recreate a clean worktree
|
|
56
|
+
git fetch origin gh-pages
|
|
57
|
+
git worktree add -B gh-pages ghp origin/gh-pages
|
|
58
|
+
if [ -f /tmp/benchmarks.tgz ]; then
|
|
59
|
+
tar -C ghp -xzf /tmp/benchmarks.tgz
|
|
60
|
+
git -C ghp add -A
|
|
61
|
+
git -C ghp commit -m "Restore benchmarks after gh-deploy" || echo "No changes"
|
|
62
|
+
git -C ghp push origin gh-pages
|
|
63
|
+
fi
|
|
64
|
+
git worktree remove ghp --force || true
|
|
@@ -5,7 +5,15 @@ all: checkstyle test test-convergence
|
|
|
5
5
|
|
|
6
6
|
# Command to run pytest for correctness tests
|
|
7
7
|
test:
|
|
8
|
-
python -m pytest --disable-warnings
|
|
8
|
+
python -m pytest --disable-warnings \
|
|
9
|
+
--cov=src/liger_kernel \
|
|
10
|
+
--cov-report=term-missing \
|
|
11
|
+
--ignore=test/convergence \
|
|
12
|
+
test/
|
|
13
|
+
|
|
14
|
+
# Command to run coverage report
|
|
15
|
+
coverage:
|
|
16
|
+
coverage report -m
|
|
9
17
|
|
|
10
18
|
# Command to run ruff for linting and formatting code
|
|
11
19
|
checkstyle:
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: liger_kernel
|
|
3
|
-
Version: 0.6.
|
|
3
|
+
Version: 0.6.4
|
|
4
4
|
Summary: Efficient Triton kernels for LLM Training
|
|
5
5
|
License: BSD 2-CLAUSE LICENSE
|
|
6
6
|
Copyright 2024 LinkedIn Corporation
|
|
@@ -35,15 +35,14 @@ Requires-Dist: triton>=2.3.1
|
|
|
35
35
|
Provides-Extra: dev
|
|
36
36
|
Requires-Dist: transformers>=4.49.0; extra == "dev"
|
|
37
37
|
Requires-Dist: matplotlib>=3.7.2; extra == "dev"
|
|
38
|
-
Requires-Dist:
|
|
39
|
-
Requires-Dist: black>=24.4.2; extra == "dev"
|
|
40
|
-
Requires-Dist: isort>=5.13.2; extra == "dev"
|
|
38
|
+
Requires-Dist: ruff>=0.12.0; extra == "dev"
|
|
41
39
|
Requires-Dist: pytest>=7.1.2; extra == "dev"
|
|
42
40
|
Requires-Dist: pytest-xdist; extra == "dev"
|
|
41
|
+
Requires-Dist: pytest-cov; extra == "dev"
|
|
42
|
+
Requires-Dist: pytest-asyncio; extra == "dev"
|
|
43
43
|
Requires-Dist: pytest-rerunfailures; extra == "dev"
|
|
44
44
|
Requires-Dist: datasets>=2.19.2; extra == "dev"
|
|
45
45
|
Requires-Dist: seaborn; extra == "dev"
|
|
46
|
-
Requires-Dist: mkdocs; extra == "dev"
|
|
47
46
|
Requires-Dist: mkdocs-material; extra == "dev"
|
|
48
47
|
Requires-Dist: torchvision>=0.20; extra == "dev"
|
|
49
48
|
Dynamic: license-file
|
|
@@ -181,8 +180,8 @@ y = orpo_loss(lm_head.weight, x, target)
|
|
|
181
180
|
- `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)
|
|
182
181
|
|
|
183
182
|
```bash
|
|
184
|
-
|
|
185
|
-
|
|
183
|
+
pip install -e .[dev]
|
|
184
|
+
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
|
|
186
185
|
```
|
|
187
186
|
|
|
188
187
|
### Optional Dependencies
|
|
@@ -216,6 +215,9 @@ pip install -e .
|
|
|
216
215
|
|
|
217
216
|
# Setup Development Dependencies
|
|
218
217
|
pip install -e ".[dev]"
|
|
218
|
+
|
|
219
|
+
# NOTE -> For AMD users only
|
|
220
|
+
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
|
|
219
221
|
```
|
|
220
222
|
|
|
221
223
|
|
|
@@ -311,7 +313,11 @@ loss.backward()
|
|
|
311
313
|
| Phi3 & Phi3.5 | `liger_kernel.transformers.apply_liger_kernel_to_phi3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
312
314
|
| Granite 3.0 & 3.1 | `liger_kernel.transformers.apply_liger_kernel_to_granite` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
|
|
313
315
|
| OLMo2 | `liger_kernel.transformers.apply_liger_kernel_to_olmo2` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
316
|
+
| Olmo3 | `liger_kernel.transformers.apply_liger_kernel_to_olmo3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
314
317
|
| GLM-4 | `liger_kernel.transformers.apply_liger_kernel_to_glm4` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
318
|
+
| InternVL3 | `liger_kernel.transformers.apply_liger_kernel_to_internvl` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
319
|
+
| HunyuanV1 | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
320
|
+
| HunyuanV1 MoE | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
315
321
|
|
|
316
322
|
|
|
317
323
|
## Low-level APIs
|
|
@@ -391,17 +397,17 @@ loss.backward()
|
|
|
391
397
|
<td style="padding: 10px;">
|
|
392
398
|
<div style="display: block;">
|
|
393
399
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
|
|
394
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=
|
|
400
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
395
401
|
</a>
|
|
396
402
|
</div>
|
|
397
403
|
<div style="display: block;">
|
|
398
404
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
|
|
399
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=
|
|
405
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
400
406
|
</a>
|
|
401
407
|
</div>
|
|
402
408
|
<div style="display: block;">
|
|
403
409
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml">
|
|
404
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=
|
|
410
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
405
411
|
</a>
|
|
406
412
|
</div>
|
|
407
413
|
</td>
|
|
@@ -129,8 +129,8 @@ y = orpo_loss(lm_head.weight, x, target)
|
|
|
129
129
|
- `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)
|
|
130
130
|
|
|
131
131
|
```bash
|
|
132
|
-
|
|
133
|
-
|
|
132
|
+
pip install -e .[dev]
|
|
133
|
+
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
|
|
134
134
|
```
|
|
135
135
|
|
|
136
136
|
### Optional Dependencies
|
|
@@ -164,6 +164,9 @@ pip install -e .
|
|
|
164
164
|
|
|
165
165
|
# Setup Development Dependencies
|
|
166
166
|
pip install -e ".[dev]"
|
|
167
|
+
|
|
168
|
+
# NOTE -> For AMD users only
|
|
169
|
+
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
|
|
167
170
|
```
|
|
168
171
|
|
|
169
172
|
|
|
@@ -259,7 +262,11 @@ loss.backward()
|
|
|
259
262
|
| Phi3 & Phi3.5 | `liger_kernel.transformers.apply_liger_kernel_to_phi3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
260
263
|
| Granite 3.0 & 3.1 | `liger_kernel.transformers.apply_liger_kernel_to_granite` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
|
|
261
264
|
| OLMo2 | `liger_kernel.transformers.apply_liger_kernel_to_olmo2` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
265
|
+
| Olmo3 | `liger_kernel.transformers.apply_liger_kernel_to_olmo3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
262
266
|
| GLM-4 | `liger_kernel.transformers.apply_liger_kernel_to_glm4` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
267
|
+
| InternVL3 | `liger_kernel.transformers.apply_liger_kernel_to_internvl` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
268
|
+
| HunyuanV1 | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
269
|
+
| HunyuanV1 MoE | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
|
|
263
270
|
|
|
264
271
|
|
|
265
272
|
## Low-level APIs
|
|
@@ -339,17 +346,17 @@ loss.backward()
|
|
|
339
346
|
<td style="padding: 10px;">
|
|
340
347
|
<div style="display: block;">
|
|
341
348
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
|
|
342
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=
|
|
349
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
343
350
|
</a>
|
|
344
351
|
</div>
|
|
345
352
|
<div style="display: block;">
|
|
346
353
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
|
|
347
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=
|
|
354
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
348
355
|
</a>
|
|
349
356
|
</div>
|
|
350
357
|
<div style="display: block;">
|
|
351
358
|
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml">
|
|
352
|
-
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=
|
|
359
|
+
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?branch=main&event=push" alt="Build">
|
|
353
360
|
</a>
|
|
354
361
|
</div>
|
|
355
362
|
</td>
|