ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1
2002 commits 1 branch 0 tags 37 MiB
Commit graph

358 commits

Author SHA1 Message Date
Ziteng Zhang
986f65cea9 [LLM] Add trust_remote_code for local renamed model in bigdl_llm_model.py (#9762) 2023-12-25 11:31:14 +08:00
Guancheng Fu
daf536fb2d vLLM: Apply attention optimizations for selective batching (#9758) 
* fuse_rope for prefil

* apply kv_cache optimizations

* apply fast_decoding_path

* Re-enable kv_cache optimizations for prefill

* reduce KV_CACHE_ALLOC_BLOCK for selective_batching
 2023-12-25 10:29:31 +08:00
Qiyuan Gong
4c487313f2 Revert "[LLM] IPEX auto importer turn on by default for XPU (#9730)" (#9759) 
This reverts commit 0284801fbd.
 2023-12-22 16:38:24 +08:00
Qiyuan Gong
0284801fbd [LLM] IPEX auto importer turn on by default for XPU (#9730) 
* Set BIGDL_IMPORT_IPEX default to true, i.e., auto import IPEX for XPU.
* Remove import intel_extension_for_pytorch as ipex from GPU example.
* Add support for bigdl-core-xe-21.
 2023-12-22 16:20:32 +08:00
Guancheng Fu
fdf93c9267 Implement selective batching for vLLM (#9659) 
* add control to load hf model

* finish initial version of selective_batching

* temp

* finish

* Remove print statement

* fix error

* Apply yang's optimization

* a version that works

* We need to check kv_cache passed in, this could be an error. TODO: add fast decoding path

* format

* temp solution: not batching prefill requests

* a version that works for prefill batching

* format

* a solid version: works normally

* a temp version

* Solid version: remove redundant functions

* fix format

* format

* solid: add option to enable selective_batching

* remove logic for using transformer models

* format

* format

* solid: enable argument VLLM_ENABLE_SELECTIVE_BATCHING

* format

* finish

* format
 2023-12-22 13:45:46 +08:00
Ruonan Wang
2f36769208 LLM: bigdl-llm lora support & lora example (#9740) 
* lora support and single card example

* support multi-card, refactor code

* fix model id and style

* remove torch patch, add two new class for bf16, update example

* fix style

* change to training_mode

* small fix

* add more info in help

* fixstyle, update readme

* fix ut

* fix ut

* Handling compatibility issues with default LoraConfig
 2023-12-22 11:05:39 +08:00
SONG Ge
ba0b939579 [LLM] Support transformers-v4.36.0 on mistral model (#9744) 
* add support transformers-v4.36.0 on mistral model

* python/llm/src/bigdl/llm/transformers/models/mistral.py

* make the redundant implementation as utils

* fix code style

* fix

* fix style

* update with utils enough_kv_room
 2023-12-22 09:59:27 +08:00
Xin Qiu
e36111e713 mixstral fused qkv and rope (#9724) 
* mixstral fused qkv and rope

* fix and clean

* fix style

* update

* update

* fix

* update

* fix
 2023-12-22 09:26:35 +08:00
Jiao Wang
e4f6e43675 safetenor to false (#9728) 2023-12-21 14:41:51 -08:00
Yishuo Wang
426660b88e simplify qwen attention (#9747) 2023-12-21 17:53:29 +08:00
Wang, Jian4
984697afe2 LLM: Add bloom gguf support (#9734) 
* init

* update bloom add merges

* update

* update readme

* update for llama error

* update
 2023-12-21 14:06:25 +08:00
Heyang Sun
df775cf316 fix python style (#9742) 
* fix python style

* fix

* fix
 2023-12-21 11:25:05 +08:00
Xin Qiu
6c3e698bf1 mistral decoding_fast_path and fused mlp (#9714) 
* mistral decoding_fast_path and fused mlp

* meet code review
 2023-12-21 10:11:37 +08:00
Heyang Sun
d157f623b6 Load Mixtral gguf in a block-wise way (#9725) 
* Load Mixtral gguf in a block-wise way

* refine
 2023-12-21 10:03:23 +08:00
Zhao Changmin
4bda975a3e LLM: Align lowbit model config (#9735) 
* align lowbit model config
 2023-12-21 09:48:58 +08:00
Wang, Jian4
e1e921f425 LLM: gguf other model using dtype (#9729) 2023-12-21 09:33:40 +08:00
Yishuo Wang
13ea6330bd optimize qwen rope (#9737) 2023-12-20 17:34:34 +08:00
Ziteng Zhang
4c032a433e [LLM] Add glibc checker (#9624) 
* Add glibc checker
* Add env BIGDL_GLIBC_CHECK to control glibc checker. The default is false, i.e., don't check.
 2023-12-20 16:52:43 +08:00
Yina Chen
cd652a1710 Support fp8 e5m2 on arc (#9711) 
* init

* fix style

* update

* fix style

* update
 2023-12-20 16:26:17 +08:00
Yishuo Wang
e54c428d30 add bf16/fp16 fuse mlp support (#9726) 2023-12-20 10:40:45 +08:00
Heyang Sun
612651cb5d fix typo (#9723) 2023-12-20 09:41:59 +08:00
Yishuo Wang
522cf5ed82 [LLM] Improve chatglm2/3 rest token performance with long context (#9716) 2023-12-19 17:29:38 +08:00
Yishuo Wang
f2e6abb563 fix mlp batch size check (#9718) 2023-12-19 14:22:22 +08:00
Heyang Sun
1fa7793fc0 Load Mixtral GGUF Model (#9690) 
* Load Mixtral GGUF Model

* refactor

* fix empty tensor when to cpu

* update gpu and cpu readmes

* add dtype when set tensor into module
 2023-12-19 13:54:38 +08:00
Qiyuan Gong
d0a3095b97 [LLM] IPEX auto importer (#9706) 
* IPEX auto importer and get_ipex_version.
* Add BIGDL_IMPORT_IPEX to control auto import, default is false.
 2023-12-19 13:39:38 +08:00
Yang Wang
f4fb58d99c fusing qkv project and rope (#9612) 
* Try fusing qkv project and rope

* add fused mlp

* fuse append cache

* fix style and clean up code

* clean up
 2023-12-18 16:45:00 -08:00
Cengguang Zhang
4d22add4af LLM: fix qwen efficiency issue in perf-test. 2023-12-18 18:32:54 +08:00
Ruonan Wang
8ed89557e5 LLM: add mlp optimization of mixtral (#9709) 2023-12-18 16:59:52 +08:00
Xin Qiu
320110d158 handle empty fused norm result (#9688) 
* handle empty fused norm result

* remove fast_rms_norm

* fix style
 2023-12-18 09:56:11 +08:00
SONG Ge
d5b81af7bd Support mixtral attention optimization on transformers-v4.36.0 (#9674) 
* add example code to support mistral/mixtral attention on transformers v4.36.0

* update

* style fix

* add update for seen-tokens

* support mixtral

* rm mistral change

* small fix

* add more comments and remove use_cache part

---------

Co-authored-by: plusbang <binbin1.deng@intel.com>
 2023-12-15 14:30:23 +08:00
Cengguang Zhang
adbef56001 LLM: update qwen attention forward. (#9695) 
* feat: update qwen attention forward.

* fix: style.
 2023-12-15 14:06:15 +08:00
Wang, Jian4
b8437a1c1e LLM: Add gguf mistral model support (#9691) 
* add mistral support

* need to upgrade transformers version

* update
 2023-12-15 13:37:39 +08:00
Wang, Jian4
496bb2e845 LLM: Support load BaiChuan model family gguf model (#9685) 
* support baichuan model family gguf model

* update gguf generate.py

* add verify models

* add support model_family

* update

* update style

* update type

* update readme

* update

* remove support model_family
 2023-12-15 13:34:33 +08:00
Yishuo Wang
9a330bfc2b fix fuse mlp when using q5_0 or fp8 (#9689) 2023-12-14 16:16:05 +08:00
Xin Qiu
5e46e0e5af fix baichuan2-7b 1st token performance regression on xpu (#9683) 
* fix baichuan2-7b 1st token performance regression

* add comments

* fix style
 2023-12-14 09:58:32 +08:00
Yishuo Wang
09ca540f9b use fuse mlp in qwen (#9672) 2023-12-13 17:20:08 +08:00
Ruonan Wang
c7741c4e84 LLM: update moe block convert to optimize rest token latency of Mixtral (#9669) 
* update moe block convert

* further accelerate final_hidden_states

* fix style

* fix style
 2023-12-13 16:17:06 +08:00
Xiangyu Tian
1c6499e880 [LLM] vLLM: Support Mixtral Model (#9670) 
Add Mixtral support for BigDL vLLM.
 2023-12-13 14:44:47 +08:00
Ruonan Wang
dc5b1d7e9d LLM: integrate sdp kernel for FP16 rest token inference on GPU [DG2/ATSM] (#9633) 
* integrate sdp

* update api

* fix style

* meet code review

* fix

* distinguish mtl from arc

* small fix
 2023-12-13 11:29:57 +08:00
Qiyuan Gong
5b0e7e308c [LLM] Add support for empty activation (#9664) 
* Add support for empty activation, e.g., [0, 4096]. Empty activation is allowed by PyTorch.
* Add comments.
 2023-12-13 11:07:45 +08:00
SONG Ge
284e7697b1 [LLM] Optimize ChatGLM2 kv_cache to support beam_search on ARC (#9579) 
* optimize kv_cache to support beam_search on Arc

* correctness test update

* fix query_length issue

* simplify implementation

* only enable the optimization on gpu device

* limit the beam_search support only enabled with gpu device and batch_size > 1

* add comments for beam_search case and revert ut change

* meet comments

* add more comments to describe the differece between multi-cases
 2023-12-13 11:02:14 +08:00
Ziteng Zhang
8931f2eb62 [LLM] Fix transformer qwen size mismatch and rename causal_mask (#9655) 
* Fix size mismatching caused by context_layer
* Change registered_causal_mask to causal_mask
 2023-12-12 20:57:40 +08:00
binbin Deng
59ce86d292 LLM: support optimize_model=True for Mixtral model (#9657) 2023-12-12 16:41:26 +08:00
Heyang Sun
9f02f96160 [LLM] support for Yi AWQ model (#9648) 2023-12-11 14:07:34 +08:00
Xin Qiu
82255f9726 Enable fused layernorm (#9614) 
* bloom layernorm

* fix

* layernorm

* fix

* fix

* fix

* style fix

* fix

* replace nn.LayerNorm
 2023-12-11 09:26:13 +08:00
Yina Chen
70f5e7bf0d Support peft LoraConfig (#9636) 
* support peft loraconfig

* use testcase to test

* fix style

* meet comments
 2023-12-08 16:13:03 +08:00
Xin Qiu
0b6f29a7fc add fused rms norm for Yi and Qwen (#9640) 2023-12-08 16:04:38 +08:00
Xin Qiu
5636b0ba80 set new linear status (#9639) 2023-12-08 11:02:49 +08:00
Yuwen Hu
6f34978b94 [LLM] Add more performance tests for win iGPU (more in-out pairs, RWKV model) (#9626) 
* Add supports for loading rwkv models using from_pretrained api

* Temporarily enable pr tests

* Add RWKV in tests and more in-out pairs

* Add rwkv for 512 tests

* Make iterations smaller

* Change back to nightly trigger
 2023-12-07 18:55:16 +08:00
Ruonan Wang
d9b0c01de3 LLM: fix unlora module in qlora finetune (#9621) 
* fix unlora module

* split train and inference
 2023-12-07 16:32:02 +08:00
Yishuo Wang
7319f2c227 use fused mlp in baichuan2 (#9620) 2023-12-07 15:50:57 +08:00
Xiangyu Tian
deee65785c [LLM] vLLM: Delete last_kv_cache before prefilling (#9619) 
Remove last_kv_cache before prefilling to reduce peak memory usage.
 2023-12-07 11:32:33 +08:00
Xiangyu Tian
0327169b50 [LLM] vLLM: fix memory leak in prepare_kv_cache (#9616) 
Revert modification in prepare_kv_cache to fix memory leak.
 2023-12-07 10:08:18 +08:00
Xin Qiu
13d47955a8 use fused rms norm in chatglm2 and baichuan (#9613) 
* use fused rms norm in chatglm2 and baichuan

* style fix
 2023-12-07 09:21:41 +08:00
Yina Chen
404e101ded QALora example (#9551) 
* Support qa-lora

* init

* update

* update

* update

* update

* update

* update merge

* update

* fix style & update scripts

* update

* address comments

* fix typo

* fix typo

---------

Co-authored-by: Yang Wang <yang3.wang@intel.com>
 2023-12-06 15:36:21 +08:00
Guancheng Fu
6978b2c316 [VLLM] Change padding patterns for vLLM & clean code (#9609) 
* optimize

* fix minor error

* optimizations

* fix style
 2023-12-06 15:27:26 +08:00
Zheng, Yi
d154b38bf9 Add llama2 gpu low memory example (#9514) 
* Add low memory example

* Minor fixes

* Update readme.md
 2023-12-05 17:29:48 +08:00
Ziteng Zhang
65934c9f4f [LLM] Fix Qwen causal_mask and attention_mask size mismatching (#9600) 
* Fix #9582 , caused by Qwen modified modeling_qwen.py 7f62181c94 (d2h-049182)
 2023-12-05 15:15:54 +08:00
Qiyuan Gong
f211f136b6 Configurable TORCH_LINEAR_THRESHOLD from env (#9588) 
* Add TORCH_LINEAR_THRESHOLD from env (BIGDL_LLM_LINEAR_THRESHOLD)
* Change default to 512
 2023-12-05 13:19:47 +08:00
Xiangyu Tian
5c03651309 [LLM] vLLM: Add Preempt for scheduler (#9568) 
Implement Preempt_by_recompute method for vllm.
 2023-12-03 20:16:25 +08:00
Xin Qiu
69c49d21f5 use fused rms norm (#9572) 
* use fused rms norm

* meet code review
 2023-11-30 21:47:41 +08:00
Yishuo Wang
7f6465518a support loading llama tokenizer from gguf model (#9565) 2023-11-30 14:56:12 +08:00
Yuwen Hu
34503efa6a Fix cpu pinned embedding (#9556) 2023-11-29 18:27:56 +08:00
binbin Deng
4ff2ca9d0d LLM: fix loss error on Arc (#9550) 2023-11-29 15:16:18 +08:00
Yishuo Wang
65121c7997 support loading q4_1/q5_0/q5_1/q8_0 gguf model (#9546) 2023-11-29 14:40:37 +08:00
Yuwen Hu
5f5ca38b74 [LLM Doc] Fix api doc rendering error (#9542) 
* Fix api rendering error

* Fix python style
 2023-11-29 09:17:09 +08:00
Yishuo Wang
a86c6e0b56 [LLM] support loading gguf model (#9544) 2023-11-28 15:51:15 +08:00
Xiangyu Tian
916c338772 fix bugs in vllm length check (#9543) 2023-11-28 11:09:54 +08:00
Zhao Changmin
e7e0cd3b5e CPU Pinned embedding Layer (#9538) 
* CPU Pinned embedding
 2023-11-28 09:46:31 +08:00
Guancheng Fu
963a5c8d79 Add vLLM-XPU version's README/examples (#9536) 
* test

* test

* fix last kv cache

* add xpu readme

* remove numactl for xpu example

* fix link error

* update max_num_batched_tokens logic

* add explaination

* add xpu environement version requirement

* refine gpu memory

* fix

* fix style
 2023-11-28 09:44:03 +08:00
Guancheng Fu
b6c3520748 Remove xformers from vLLM-CPU (#9535) 2023-11-27 11:21:25 +08:00
binbin Deng
6bec0faea5 LLM: support Mistral AWQ models (#9520) 2023-11-24 16:20:22 +08:00
Ruonan Wang
914a5a5a27 LLM: fix abnormal Mistral GPU accuracy by updating rms_norm (#9529) 2023-11-24 15:37:50 +08:00
SONG Ge
3d24823cda hot-fix mistral kv_cache (#9528) 2023-11-24 14:33:04 +08:00
Zhao Changmin
42b7a16bc5 Replace torch.bmm with safe_bmm (#9519) 
* replace bmm with safe one

* rename args and deprecated warning
 2023-11-24 12:16:48 +08:00
Ruonan Wang
b63aae8a8e LLM: add flash attention support for llama (#9518) 
* add initial flash attention for llama

* accelerate fp32 first token by changing to fp16 in advance

* support fp32
 2023-11-23 18:40:18 +08:00
Guancheng Fu
bf579507c2 Integrate vllm (#9310) 
* done

* Rename structure

* add models

* Add structure/sampling_params,sequence

* add input_metadata

* add outputs

* Add policy,logger

* add and update

* add parallelconfig back

* core/scheduler.py

* Add llm_engine.py

* Add async_llm_engine.py

* Add tested entrypoint

* fix minor error

* Fix everything

* fix kv cache view

* fix

* fix

* fix

* format&refine

* remove logger from repo

* try to add token latency

* remove logger

* Refine config.py

* finish worker.py

* delete utils.py

* add license

* refine

* refine sequence.py

* remove sampling_params.py

* finish

* add license

* format

* add license

* refine

* refine

* Refine line too long

* remove exception

* so dumb style-check

* refine

* refine

* refine

* refine

* refine

* refine

* add README

* refine README

* add warning instead error

* fix padding

* add license

* format

* format

* format fix

* Refine vllm dependency (#1)

vllm dependency clear

* fix licence

* fix format

* fix format

* fix

* adapt LLM engine

* fix

* add license

* fix format

* fix

* Moving README.md to the correct position

* Fix readme.md

* done

* guide for adding models

* fix

* Fix README.md

* Add new model readme

* remove ray-logic

* refactor arg_utils.py

* remove distributed_init_method logic

* refactor entrypoints

* refactor input_metadata

* refactor model_loader

* refactor utils.py

* refactor models

* fix api server

* remove vllm.stucture

* revert by txy 1120

* remove utils

* format

* fix license

* add bigdl model

* Refer to a specfic commit

* Change code base

* add comments

* add async_llm_engine comment

* refine

* formatted

* add worker comments

* add comments

* add comments

* fix style

* add changes

---------

Co-authored-by: xiangyuT <xiangyu.tian@intel.com>
Co-authored-by: Xiangyu Tian <109123695+xiangyuT@users.noreply.github.com>
Co-authored-by: leonardozcm <leonardo1997zcm@gmail.com>
 2023-11-23 16:46:45 +08:00
Qiyuan Gong
0f0c6bb631 [LLM] Fix Qwen registered_causal_mask is None (#9513) 
* Add registered_causal_mask init based on 2abd8e5777.
 2023-11-23 09:28:04 +08:00
Ruonan Wang
076d106ef5 LLM: GPU QLoRA update to bf16 to accelerate gradient checkpointing (#9499) 
* update to bf16 to accelerate gradient checkpoint

* add utils and fix ut
 2023-11-21 17:08:36 +08:00
Xin Qiu
50b01058f1 enable new q4_1 (#9479) 2023-11-17 14:58:57 +08:00
Zhao Changmin
30abd304a7 LLM: Fix baichuan pre-normalize model tensor assigning issue when loading (#9481) 
* No need to normalized when loading
 2023-11-16 21:57:28 +08:00
Ruonan Wang
c0ef70df02 llm: quick fix of fast_rms_norm (#9480) 2023-11-16 14:42:16 +08:00
Yina Chen
d5263e6681 Add awq load support (#9453) 
* Support directly loading GPTQ models from huggingface

* fix style

* fix tests

* change example structure

* address comments

* fix style

* init

* address comments

* add examples

* fix style

* fix style

* fix style

* fix style

* update

* remove

* meet comments

* fix style

---------

Co-authored-by: Yang Wang <yang3.wang@intel.com>
 2023-11-16 14:06:25 +08:00
Ruonan Wang
d2c064124a LLM: update rms related usage to suport ipex 2.1 new api (#9466) 
* update rms related usage

* fix style
 2023-11-16 11:21:50 +08:00
Yuwen Hu
731b0aaade Empty cache after embedding to cpu (#9477) 2023-11-16 10:52:30 +08:00
Yang Wang
51d07a9fd8 Support directly loading gptq models from huggingface (#9391) 
* Support directly loading GPTQ models from huggingface

* fix style

* fix tests

* change example structure

* address comments

* fix style

* address comments
 2023-11-13 20:48:12 -08:00
SONG Ge
2888818b3a [LLM] Support mixed_fp8 on Arc (#9415) 
* ut gpu allocation memory fix

* support mix_8bit on arc

* rename mixed_4bit to mixed_fp4 and mixed_8bit to mixed_fp8

* revert unexpected changes

* revert unexpected changes

* unify common logits

* rename in llm xmx_checker

* fix typo error and re-unify
 2023-11-13 09:26:30 +08:00
Heyang Sun
df8e4d7889 [LLM] apply allreduce and bias to training in LowBitLinear (#9395) 2023-11-09 14:35:54 +08:00
Wang, Jian4
40cead6b5b LLM: Fix CPU qlora dtype convert issue (#9394) 2023-11-09 14:34:01 +08:00
Ruonan Wang
bfca76dfa7 LLM: optimize QLoRA by updating lora convert logic (#9372) 
* update convert logic of qlora

* update

* refactor and further improve performance

* fix style

* meet code review
 2023-11-08 17:46:49 +08:00
Ruonan Wang
7e8fb29b7c LLM: optimize QLoRA by reducing convert time (#9370) 2023-11-08 13:14:34 +08:00
Yishuo Wang
bfd9f88f0d [LLM] Use fp32 as dtype when batch_size <=8 and qtype is q4_0/q8_0/fp8 (#9365) 2023-11-08 09:54:53 +08:00
Heyang Sun
fae6db3ddc [LLM] refactor cpu low-bit forward logic (#9366) 
* [LLM] refactor cpu low-bit forward logic

* fix style

* Update low_bit_linear.py

* Update low_bit_linear.py

* refine
 2023-11-07 15:09:16 +08:00
Heyang Sun
af94058203 [LLM] Support CPU deepspeed distributed inference (#9259) 
* [LLM] Support CPU Deepspeed distributed inference

* Update run_deepspeed.py

* Rename

* fix style

* add new codes

* refine

* remove annotated codes

* refine

* Update README.md

* refine doc and example code
 2023-11-06 17:56:42 +08:00
Xin Qiu
1420e45cc0 Chatglm2 rope optimization on xpu (#9350) 2023-11-06 13:56:34 +08:00
Yuwen Hu
a0150bb205 [LLM] Move embedding layer to CPU for iGPU inference (#9343) 
* Move embedding layer to CPU for iGPU llm inference

* Empty cache after to cpu

* Remove empty cache as it seems to have some negative effect to first token
 2023-11-03 11:13:45 +08:00
Yishuo Wang
726203d778 [LLM] Replace Embedding layer to fix it on CPU (#9254) 2023-11-01 13:58:10 +08:00
Yang Wang
e1bc18f8eb fix import ipex problem (#9323) 
* fix import ipex problem

* fix style
 2023-10-31 20:31:34 -07:00
Yina Chen
2262ae4d13 Support MoFQ4 on arc (#9301) 
* init

* update

* fix style

* fix style

* fix style

* meet comments
 2023-11-01 10:59:46 +08:00
Yang Wang
163d033616 Support qlora in CPU (#9233) 
* support qlora in CPU

* revert example

* fix style
 2023-10-27 14:01:15 -07:00