Qwen1.5-4b and Qwen1.5-7b model cannot be loaded correctly in ipex-llm version 20240522 about bigdl HOT 9 OPEN

we have made some breaking change on qwen-1.5's int4 checkpoint in 5.21 version, old int4 checkpoint(generated by ipex 0520 or eariler) cannot be loaded with new ipex-llm(0521 or later), please regenerate int4 checkpoint with ipex-llm 20240521 or later

from bigdl.

we have made some breaking change on qwen-1.5's int4 checkpoint in 5.21 version, old int4 checkpoint(generated by ipex 0520 or eariler) cannot be loaded with new ipex-llm(0521 or later), please regenerate int4 checkpoint with ipex-llm 20240521 or later

ok, got it.
the new version has some improvements? such as quantization accuracy, or RAM?

from bigdl.

ok, got it.
the new version has some improvements? such as quantization accuracy, or RAM?

yes, there should be some improvements on speed and RAM, but not much

from bigdl.

ok, got it.
the new version has some improvements? such as quantization accuracy, or RAM?

yes, there should be some improvements on speed and RAM, but not much

I regenerate qwen-7b int4 model and run it on my laptop(ultra 7 155H), but the "warm up" stage costs very long time(more than 5 minutes), do you have any advice?

from bigdl.

I regenerate qwen-7b int4 model and run it on my laptop(ultra 7 155H), but the "warm up" stage costs very long time(more than 5 minutes), do you have any advice?

Did you set SYCL_CACHE_PERSISTENT=1? https://bigdl.readthedocs.io/en/latest/doc/LLM/Overview/install_gpu.html#runtime-configuration

from bigdl.

I regenerate qwen-7b int4 model and run it on my laptop(ultra 7 155H), but the "warm up" stage costs very long time(more than 5 minutes), do you have any advice?

Did you set SYCL_CACHE_PERSISTENT=1? https://bigdl.readthedocs.io/en/latest/doc/LLM/Overview/install_gpu.html#runtime-configuration

yes, i have set it
I found that warm up speed is much faster in cpu mode(about 10-20s). but slower in xpu mode..

from bigdl.

I found that warm up speed is much faster in cpu mode(about 10-20s). but slower in xpu mode..

CPU doesn't need JIT compilation, while gpu needs.

On CPU: load model -> quantization -> inference

On GPU: load model -> quantization -> JIT compilation -> inference. This JIT compilation is what we called warm up, and it costs about ten minutes.

set SYCL_CACHE_PERSISTENT=1 will store gpu JIT code on disk so that it won't need to compile again the second time you run it.

If you are using powershell, please use CMD instead.

Could you check whether C:\Users\<user name>\AppData\Roaming\libsycl_cache exists ? If exits, please delete it. Then set SYCL_CACHE_PERSISTENT=1 and run inference (this run will take a long time (about 10 minutes) because it needs to regenerate JIT code cache), after finish, you should see regenerated C:\Users\<user name>\AppData\Roaming\libsycl_cache. With cache, following inference should has no warm up. (set SYCL_CACHE_PERSISTENT=1 is still required)

from bigdl.

I found that warm up speed is much faster in cpu mode(about 10-20s). but slower in xpu mode..

CPU doesn't need JIT compilation, while gpu needs.

On CPU: load model -> quantization -> inference

On GPU: load model -> quantization -> JIT compilation -> inference. This JIT compilation is what we called warm up, and it costs about ten minutes.

set SYCL_CACHE_PERSISTENT=1 will store gpu JIT code on disk so that it won't need to compile again the second time you run it.

If you are using powershell, please use CMD instead.

Could you check whether C:\Users\<user name>\AppData\Roaming\libsycl_cache exists ? If exits, please delete it. Then set SYCL_CACHE_PERSISTENT=1 and run inference (this run will take a long time (about 10 minutes) because it needs to regenerate JIT code cache), after finish, you should see regenerated C:\Users\<user name>\AppData\Roaming\libsycl_cache. With cache, following inference should has no warm up. (set SYCL_CACHE_PERSISTENT=1 is still required)

ok，i will try， thank you very much.
If libsycl_cache exists, even if I finish the infer process, restart and reload model, is there no need for a warm up?

from bigdl.

If libsycl_cache exists, even if I finish the infer process, restart and reload model, is there no need for a warm up?

yes

from bigdl.