码涯-AIGC代码仓库-openoker/DeepSpeed: DeepSpeed是一个深度学习优化库，它使分布式训练和推理变得简单、高效。 @ 13847c15b5d0c6c0bc9c27e14f0dedd0581e08d0

目录树: 13847c15b5

AutoPR/0.12.2

AutoPR/0.14.0

AutoPR/0.14.5

CUDA-Graph-support

HeyangQin/deepspeed-ulysses-chinese-blog

HeyangQin/enable_hpz_nograd

HeyangQin/fastgen_moe_h100

HeyangQin/fix_hpz_nograd

HeyangQin/fix_issue_3062

HeyangQin/fix_issue_3068

HeyangQin/fix_issue_3156

HeyangQin/fix_issue_5205

HeyangQin/fix_pr_3462_standalone

HeyangQin/hpz_convergence

HeyangQin/inference_t5_phase1

HeyangQin/mixed_precision_lora_sam

HeyangQin/mixz_tutorial

HeyangQin/skip_bias_quant

HeyangQin/staging-zero-pp-v1

HeyangQin/ucp_blog_chinese

HeyangQin/ulysses_fp8

Megtron-Kernel-Integration

SA_feature_tag

SA_tutorial_update

SA_update_tutorial_link

add-bfp16-support

add-comm-layout

add-inference-comm

add-llama2-support

add-quantizer

add-shared-lib

adk9/phi3-inference

adk9/phi3-small

adk9/update-minor-cuda

amawa/1-bit-alltoall

amawa/1bit-adam-nccl

amawa/add-moe-container

amawa/aml-get-hosts

amawa/auto-save-ckpt

amawa/config-pass-down

amawa/debug

amawa/fix-amd-rocm

amawa/fix-auto-tp-load-ckpt

amawa/fix-tracer-zero3

amawa/fix-z3-for-hf-accelerate

amawa/fix-z3-warn-print-v2

amawa/inference-fix

amawa/remove-deepcopy

amawa/split-a2a

amawa/zero-inf-refactor

amd-jiting

aml-autotuner

arashb-patch-1

arashb/fix-phi-2

arpan/auto-check

autocast-fix

awan-10-patch-1

awan-10-patch-2

awan-10-patch-3

azure

big-science

big-science-v2

bing/debugging

bing/ds-adam

bing/formatting-correction

bing/io-tutorial

bing/modify-ds-optimizer

bing/optimizer-naming

bloom-debug

chatgpt-chinese-blog

check-linear-sizes

cholmes/activation-utils

cholmes/checkpoints-inference-v2-2

cholmes/comm-group-cache

cholmes/fix-asym-quant

cholmes/fix_reduction_utils_amd

cholmes/isolate-src-code

cholmes/kv-cache-flexibility

cholmes/mem-access-predicated-load

cholmes/migrate-to-dequant-lib

cholmes/pipelined-quant

cholmes/reduce-quantized-gpus

cholmes/sd-extension

cholmes/ts-builder

cholmes/unique-cuda-graphs

ckpt-fix-unfused

clean-llama

clean-llama-v2

clean-opt

clean-opt-base

clean-opt-v2

clean-opt-v2-base

codegen-inference

comm-opt2

costineseanu/windows_inference_build

cpu-adam/optional_CUDA-copy

debug-base-attn

debug-ds-inf

debug-ds-inf-torch-matmul

ds-chat-blog-8-31

ds-chat-clean-opt

ds-chat-news

ds-chat-release

ds-inference/add-falcon-support

ds-inference/bloom-support-meta

ds-inference/fix-generation

ds-inference/fix-mp

ds-inference/remove-randgen

ds-inference/simplify

ds-inference/support-large-token-length

ds-seq-tutorial

ds-vchat-blog-v1

ds-vchat-blog-v2

duli/capability

duli/cuda_op_builder

duli/op_builder

duli/pre_post

duli/zero_debugging

elastic-ckpt-refresh

elasticity-v2

eltonz/copy_grad_stream

enable-neox

encoded-ds-config

fairseq-moe

fairseq-moe-debug

falcon-180b

fastgen-blog

fastgen-blog-2

features/rebase-quant-fp6

fix-MoQ

fix-autotuning-docs

fix-autotuning-exit

fix-autotuning-reqs

fix-flops-profiler

fix-fp16-test

fix-injection

fix-max_train_batch_size

fix-misaligned-grad

fix-moe-top1gating

fix-sp-dense

fix-sparse-attn

fix-tuner-prescale_gradients

fix-tuner-scheduler-bug

fix-twitter

fix-typos

fix_mpu_ckpt

flash-attention

flops-profiler-skip-unused-args

fp6-blog

fs-82

fs-soft-kernel

fs-z2-fix

fs/soft-kernel

gcooper/make_optimizer_optional

generic-ckpt-loading

gh-pages

gh-readonly-queue/master/pr-3852-3491e32d72746ec3d990108a23e67b2666b3e0e0

gh-readonly-queue/master/pr-3852-adb9bc14b780115fd54f3f1234abcb7ab52fa975

gh-readonly-queue/master/pr-3854-85503dab878875175b6d5eb6a39125878c172273

gh-readonly-queue/master/pr-3892-548451ba4e8ea71029d738c33f639e0439aad1dd

gh-readonly-queue/master/pr-3892-9f8817b2425bb82d9b6355caa6d2d0ebd036885d

gh-readonly-queue/master/pr-3893-cc71eec8c85c4437d8139e53372da7f22224fed5

gh-readonly-queue/master/pr-3928-82115d9059ce8271229c8f63153a02f2d323cfc1

gh-readonly-queue/master/pr-4163-5e16eb2c939707d0d0062a458d77998fccb3afad

gma/xpu_compile_analysis

good-moe

gpt2-debug

guanhua/adam-timer

guanhua/adam-timer2

guanhua/check-bf16

guanhua/fix-cutlass-ver

guanhua/h2d-offload

guanhua/kernel-test

guanhua/mics-fix

guanhua/overflow-check

guanhua/quant-dequant-test

guanhua/quant-test

guanhua/rocm-cpu-adam

guanhua/v14.0-bf16-check

hf-workaround

hp-sam

hpzero-preview

inference-api/tutorial

inference-read-checkpoint

inference-refactor-v1-mro-test

inference/ElutherAI-GPTJ

inference/TP-general-support

inference/add-bf16-support

inference/engine-api

inference/fix-masking

inference/fix-mp-init

inference/support-encoder-decoder

injection-fixes

jeff-test

jeffra-patch-2

jeffra/1node-launcher-fix

jeffra/2904

jeffra/auto-bucket

jeffra/available_memory

jeffra/bf16-updates

jeffra/bf16-updates-v2

jeffra/ci-updates

jeffra/ckpt-barrier

jeffra/docker-update

jeffra/engine-xthru

jeffra/engine-xthru-v2

jeffra/engine-xthru-v2-no-padding

jeffra/external-skip

jeffra/fix-1416

jeffra/fs-diverge

jeffra/fs-gas-fix

jeffra/fs-gas-fix-v2

jeffra/fs-support

jeffra/fs-z3

jeffra/fs-z3-v0510

jeffra/gptj-fixes

jeffra/inf-engine-refactor

jeffra/inf-tests

jeffra/jit-fix

jeffra/latest-hf

jeffra/op-build-api

jeffra/prepost_fwd_and_generate

jeffra/saksham-zero1-fixes

jeffra/savepid2

jeffra/shm-report

jeffra/staging-comms-logging-v1

jeffra/turn-on-opt-test

jeffra/update-z3-check

jeffra/z1-refresh

jeffra/z1-refresh-2

jeffra/z1-refresh-3

jeffra/z3-fix

jeffra/z3-new-param

jeffra/zero-1-fix

jeffra/zero-1-fix-test

jeffra/zero-ckpt-fixes

jeffra/zero-moe-noCG

jeffra/zero1-grad-norm

jerasley/mac

jomayeri/aio-file-offset

jomayeri/aio-locked-tensor

jomayeri/aio-mem-fix

jomayeri/aio-type-mismatch

jomayeri/bf16-zero-check

jomayeri/bug-5880

jomayeri/debug-2361

jomayeri/deepnvme-perf-debug

jomayeri/destroy-zero

jomayeri/fp8-init

jomayeri/gds-swapper-fix

jomayeri/h100-unittest

jomayeri/he-mp-assert

jomayeri/issue-3367

jomayeri/issue-3560

jomayeri/issue-3598

jomayeri/issue-3769

jomayeri/issue-4083

jomayeri/issue-4095

jomayeri/issue-4183

jomayeri/issue-5087

jomayeri/lr-step-init

jomayeri/lr-step-move

jomayeri/model-param-list

jomayeri/new-zero-accum

jomayeri/swap-with-locked

jomayeri/zero-grad-accum

jomayeri/zero3-hooks

kv-cache-reset

landing-training

landing-updates

lekurile/add_ds_chat_workflow

lekurile/add_hip_abstraction

lekurile/clean_up_params

lekurile/container_param_cleanup

lekurile/ds_chat_attn_mlp_base

lekurile/ds_chat_fix_test

lekurile/ds_chat_gh_wf

lekurile/ds_chat_mlp_debug

lekurile/ds_chat_revert_54c06872

lekurile/ds_chat_test_54c06872

lekurile/ds_chat_test_7b5b0660

lekurile/ds_chat_test_exit_first

lekurile/ds_chat_test_f69f8840

lekurile/fix_ds_chat_bloom

lekurile/fix_formatting

lekurile/fix_he_print

lekurile/fix_issue_2330

lekurile/fix_opt_meta_tensor

lekurile/fix_phi_2

lekurile/fix_sd

lekurile/fix_sd_ci

lekurile/fix_unet_vae

lekurile/general_local_cg

lekurile/infv2_lm_eval

lekurile/kernel_hip_amd

lekurile/load_ckpt_inf_eng

lekurile/mlp_functions

lekurile/offload_fix_test

lekurile/sd_min_ver

lekurile/test_rearrange_ops

lekurile/update_ds_chat_ci

lekurile/update_ds_chat_ci_2

lekurile/update_ds_chat_ci_test

lekurile/update_dschat_wf

lekurile/update_inf_ckpt_load

lf-test

loadams/adam-params

loadams/add-contributing-release-md-files

loadams/add-gaudi-badge-readme

loadams/add-scheduled-open-issue-check-ds-chat

loadams/add-torch-2-support

loadams/amd-57

loadams/amd-mi200-tests

loadams/amd-pre-compile

loadams/amd-updates

loadams/auto-stage3-prefetch-bucket-size

loadams/auto-task-open-failure

loadams/build-for-cpu

loadams/changes-to-op-builder

loadams/check-pydantic-v2-support

loadams/clear-cache

loadams/cpu-inf

loadams/cpu-inf-triggers

loadams/cpu-inf-v0-docker

loadams/cpu-inference-shorten

loadams/cpu-torch

loadams/cu118

loadams/debug-torch

loadams/disable-h100-ci

loadams/disable-windows-ops-build-script

loadams/dot-deepspeed_env-test

loadams/dpkg-libaio

loadams/empty-env-var-setup

loadams/enable-amdmi200

loadams/enable-workflow-dispatch-nv-torch-nightly-v100

loadams/engine-pos-args

loadams/fix-check-valid-version

loadams/fix-cpu-inf-test-time

loadams/fix-cuda-build-ops

loadams/fix-fp16-bf16-logging-issue

loadams/fix-hpu

loadams/fix-lightning-pytorch2

loadams/fix-mpi4py

loadams/fix-nccl-comm-torch-check

loadams/fix-no-torch-failure-mlu

loadams/fix-nv-inference

loadams/fix-nv-inference-hang

loadams/fix-nv-torch-latest-v100

loadams/fix-onebit-skip

loadams/fix-torch-2

loadams/fix-torch-compiler-hasattr

loadams/fix-triggers-no-torch-workflow

loadams/flops-profiler-scaled-dor-attn-torch-2

loadams/get-amd-team-ci

loadams/gh-cpu-inf

loadams/gh-release-version-update

loadams/hf-transformers-ci-fix

loadams/hpu-uts

loadams/ignore-unused-params-default

loadams/lamb-bf16

loadams/libaio

loadams/low-cpu-mem-ut

loadams/lsb-release

loadams/megatron

loadams/megatron-lm-112

loadams/megatron-new-pypi

loadams/megatron-version

loadams/more-torch-2-support

loadams/nv-inf-jobs-test

loadams/nv-inf-test

loadams/nv-inference-revert

loadams/nv-nightly

loadams/nv-nightly-fix-transformers

loadams/nv-sd-badge

loadams/openmpi-eth0

loadams/pin-torch-latest-ver

loadams/py36

loadams/pynvml

loadams/recurse-flops-profiler

loadams/reenable-cpu-inference

loadams/remove-dead-code

loadams/remove-modeling

loadams/remove-python-36-check

loadams/rename-fp-quantize-cu

loadams/rename-nv-torch-latest-cpu-workflow

loadams/revert-4660

loadams/revert-5608

loadams/revert-cpu-inf

loadams/revert-loss

loadams/revert-nv-inference-changes

loadams/revert-pr-5608

loadams/revert-userwarning

loadams/rocm-fixes

loadams/rocm57

loadams/rocm6

loadams/sd-paths

loadams/setup-h100-triggers

loadams/sigterm

loadams/skip-nv-inference

loadams/sparse-attn-fix

loadams/sparse-attn-torch-2

loadams/stablediffusion-test-triton2

loadams/switch-modeling-compression

loadams/switch-python-versions

loadams/tar-vuln

loadams/test-0.15.0

loadams/test-amp-futurewarning

loadams/test-b421e8c8f31af254b63ad6e9839f617ab6d9c060

loadams/test-ccl-fixes

loadams/test-compile

loadams/test-cpu

loadams/test-cpu-inf-fix

loadams/test-f0e3f01d7c7a3d8748212e61eaf487fab41168a7

loadams/test-fix-nv-inference

loadams/test-glibc228

loadams/test-merged-changes

loadams/test-model-task

loadams/test-nv-ds-chat-failure-mode

loadams/test-nv-latest-cpu

loadams/test-nv-torch-latest-v100

loadams/test-pytest-ordering

loadams/test-runsc

loadams/test-torch-2.3.0

loadams/torch-cpu-mismatch-cudaopbuilder

loadams/torch-nightly-debug

loadams/transformers-fixes

loadams/transformers-torch

loadams/transformers-torch-update

loadams/transformers-workflow-dispatch

loadams/triton-22-update

loadams/try-bump-pydantic

loadams/unpin-hf-transformers-nv-workflows

loadams/unpin-nv-torch-latest

loadams/unpin-transformers

loadams/update-2004-checkout-actions

loadams/update-accelerate

loadams/update-amd-required-paths

loadams/update-checkout

loadams/update-conda-pydantic

loadams/update-container-a6000

loadams/update-docker

loadams/update-dockerfile

loadams/update-hostname-I

loadams/update-hpu-docker-container

loadams/update-hpu-docker-image

loadams/update-hpu-gaudi-flow-more

loadams/update-nodejs-reate-pr-action

loadams/update-nv-accelerate

loadams/update-nv-inference-torch-ver

loadams/update-nv-lightning-test-cu-ver

loadams/update-nv-torch-latest-cpu-torch-ver

loadams/update-nv-torch-latest-cpu-version

loadams/update-pre-compile-ops-docker

loadams/update-pydantic

loadams/update-pytest

loadams/update-pytest-error-codes

loadams/update-real-latest

loadams/update-sd-triton

loadams/update-torch-113

loadams/update-transformers

loadams/update-transformers-cu116

loadams/update-version-txt-post-release

loadams/update-website-sidebar

loadams/x86-accelerator

loadams/xpu-readme

loadams/xpu-test

loadams/xpu-yml

lokoppak/ln_schedule_update

lokoppak/low_cpu_mem_usage_ut

lokoppak/new_pt_binding

lokoppak/quantization_3d

lokoppak/ref_ln

lsh

master

master-test

megatron2.4-3d

minjiaz/ds-seq-tutorial

minjiaz/moe-comm

minjiaz/moe-sharing

moe-full-tp

moe-inference-tutorial

moe-inference-tutorial1

moe-inference/add-tutorial

moe-pipelining

moe-timing

mosm/autotp-he

mosm/autotp_llama

mosm/bloom_dev

mosm/codegen

mosm/debug-ds-attn

mosm/debugger

mosm/dschat-news

mosm/inf-refactor

mosm/llama2

mosm/matmul_test

mosm/module_parser

mosm/mp_tutorial

mosm/opt-kernel

mosm/softmax

mosm/softmax-longseq

mosm/t5

mosm/test

mosm/tp_dev

mosm/wb-param

mrwyattii/expand-fp16-tests

mrwyattii/fix-accelerate-tests

mrwyattii/fix-for-mii-UT

mrwyattii/fix-inference-skipped-tests

mrwyattii/fix-launcher-user-args

mrwyattii/fix-multi-node-checks

mrwyattii/pin-datasets

mrwyattii/pydantic-2-support

mrwyattii/remove-symlinks

mrwyattii/rename-cpu-accelerator

mrwyattii/safetensor

mrwyattii/silence-backend-warning

mrwyattii/update-GH-permission

mrwyattii/update-MII-tests-infV2

multi-z3-prs

multi-z3-prs-r2

mz/llama-support

neox-q-int8

niumanar/gan_optimizer

offloadpp-news

olruwase/accelerator_abstraction

olruwase/adam_types

olruwase/align_rrg_rs_param_order

olruwase/all_gather_profiling

olruwase/amd_configurable_pp_rtol

olruwase/assert_unused_parameters

olruwase/b16-debugging

olruwase/bf16-updates-2

olruwase/bf16_tied_weights_reduce

olruwase/bf16_update_hp_params

olruwase/bloom-support

olruwase/bloom_176b_checkpoint_bc

olruwase/build_compat_ops

olruwase/ci_pytorch_1x

olruwase/deepnvme_abstract_class

olruwase/deepnvme_docs

olruwase/disable_prefetch_profiler

olruwase/disable_z3_prefetcher

olruwase/dnvme_docs

olruwase/ds_2449

olruwase/ds_2921

olruwase/ds_3481

olruwase/ds_3680_2

olruwase/ds_3948

olruwase/dynamic_graph_activation_checkpoint

olruwase/elastic-ckpt-refresh

olruwase/engine_destroy

olruwase/fix_kernel_memory_bloat

olruwase/frozen_weights_unit_test

olruwase/fs-zero3_trace_fix

olruwase/fs_z3_trace_error_disable

olruwase/fs_z3_trace_log

olruwase/fuse_torch_adam_w

olruwase/gpt3-finetuning

olruwase/grad_accum_loss

olruwase/issue_3062

olruwase/llama2_empty_group

olruwase/local_storage_checkpoint

olruwase/lr_warmup_decay

olruwase/non_tensor_activation_checkpoint

olruwase/nvme_finetune

olruwase/nvme_offload_bug

olruwase/nvme_perf_sweep

olruwase/nvme_testsuite

olruwase/override_module_apply

olruwase/refactor_universal_checkpoint

olruwase/restore_from_bit16_weights

olruwase/round_robin_gradient_option

olruwase/safe_pkg_check

olruwase/safe_py_subprocess

olruwase/save_checkpoint_latest_false

olruwase/save_zero3_fp16_weights

olruwase/set_zero_opt_grad

olruwase/setup_env_libaio

olruwase/trainable_parameters

olruwase/windows_blog

olruwase/z3_perf_tune

olruwase/z3_suppress_warning

olruwase/zcode_model_expert

olruwase/zero1_non_tensor_checkpoint

olruwase/zero2_grad_accum_bug

olruwase/zero2_offload_keyerror

olruwase/zero2_offload_rrb_divergence

olruwase/zero2_offload_slowdown

olruwase/zero2_trainable_parameters

olruwase/zero2_trainable_parameters_v0.5.7

olruwase/zero2_unbalanced_grad_reduction

olruwase/zero3_amp_autocast

olruwase/zero3_broken_tracing

olruwase/zero3_dp_norm_allreduce

olruwase/zero3_profile_fetch

olruwase/zero3_unboundlocal_bug

olruwase/zero_inference_tokgen

olruwase/zero_inference_torch_version

olruwase/zero_multi_models

olruwase/zero_offload_e2e

olruwase/zero_offload_fix_corner_case

olruwase/zero_offload_v3

olruwase/zero_optional_reduce_scatter

olruwase/zero_stage1_checkpoint_layout

olruwase/zero_stage1_elastic_checkpoint

olruwase/zinf_none_swapper

paper

patch-z1-cont-grad

pr_moe_tutorial

preserve-CVDs

profiler-add-shape

qanthony/bigbird

qanthony/comms-bench

qanthony/nccl-backend

quantization-refresh

quantize-inference

refine-quantizer

remotes/origin/dev/tput

remove-tbx

remove-unused-quantize-settings

reyazda/adam-scalar-fix

reyazda/cpu_adam_jit_v2

reyazda/fix-inference-api

reyazda/pytorch-workspace-allocate

reyazda/remove_bertid

reyazda/support_AVX2_by_default

reyazda/test-hidden-dimension

reyazda/test-sparse

reyazda/test-sparse-v2

reyazda/test-transformer

reyazda/testing_embedding

reyazda/triton-new-sparse

reza/deepspeed_adam_merge_v3

reza/fix-adam-copyfp16

reza/fix_adam_corner_case

reza/fix_adam_perf

reza/megatron_kernel_integration

rtd-staging

saksham-zero1-fixes

samyam-overlap-comm

samyamr/elasticity

samyamr/fix-for-fragmented-linear-inputs

samyamr/gpt3-finetuning

samyamr/gpt3-finetuning-mixed-precision

samyamr/stage3-alignment-fix

samyamr/zero-2-debug

security-patch

shaden/textgen

smartreply_hotfix

sp/comm-opt

sparse-attn-cuda11

sparse-attn/support-latest-triton

staging-amd

staging-amd-port

staging-amd-v2

staging-amd-v3

staging-comms-next-v2

staging-comms-v1

staging-deepnvme-gds-v1

staging-demo-feature-v0

staging-ds-chat-blog-v1

staging-ds-seq-v1

staging-inference-v2-5

staging-mii-update

staging-moe-next-v1

staging-oaas

staging-pld-v1

staging-pp

staging-test

staging-zero-dual-v2

staging-zero-dual-v3

staging-zero-dual-v5

staging-zero-inference-v1

stale-issues

styoun/triton-flash2

styoun/triton2.1

styoun/triton2.1-autotune

styoun/zero-inf-8bit-q

subprocess-test

test-ac

test-cuda-11.7

tmp

tmp-old

tohtana/add_slides_meetup_japan

tohtana/allocate_test_port

tohtana/autocast_only_floating_values

tohtana/bcast_warning_z3

tohtana/blog_win_jp

tohtana/cache_kv_requirements

tohtana/clean_all_param_coordinators

tohtana/clean_up_prefetch_param

tohtana/compile-zero

tohtana/compile_no_grad

tohtana/consistent_zero_grad

tohtana/debug_compile_backends

tohtana/debug_semaphore_leak

tohtana/file_store_for_tests

tohtana/fix-save-checkpoint-step

tohtana/fix_bf16_opt_update_hp

tohtana/fix_chkpt_alignment

tohtana/fix_sort_dp_univ_ckpt

tohtana/fix_univ_chkpt_load

tohtana/fix_zero_init_patch

tohtana/get_offload_state_api

tohtana/ignore_reuse_dist_env

tohtana/lock_hf_cache_update

tohtana/log_run_tests

tohtana/model_declaration_in_init_context

tohtana/offload_zero_buffers

tohtana/pipeline_with_compiled_module

tohtana/remove_step_on_init

tohtana/test_with_pt25

tohtana/univ_ckpt_custom_shape

tohtana/z3_multi_dtypes

token-drop

transformer-injection

transformer-kernel/support-arbitrary-hidden

transformer/fix-layer-norm

transformer/injection

transformer/large-seq-support

transformer/triangular-mask

triton-fix

ucp_blog

umchand/test_compiler

umchand/triton/bias_act

unify-benchmark-knowledge

update-flops-profiler-doc

update-flops-profiler-pool-compute

workaround-zero3

z1-offload-multigpu

z3-mem-leak

zero-ckpt-cpu-issue-v2

zhenyzhang-data

zheweiyao/quantize_update

1234567

Mixture of Experts (MoE)
====================

Layer specification
--------------------
.. autoclass:: deepspeed.moe.layer.MoE
    :members: