Fix causal_lm cpp demo for llama architecture

[sammysun0711]

This PR aim to fix causal_lm cpp demo for llama architecture

• Change directory name from casual_lm to causal_lm
• Fix causal lm cpp demo input & output mismatch for llama architecture
  • llama does not require position ids as input, only chatglm required
  • llama past kv cache as input starting from idx 2 instead of 3

This is a initial step to build baseline for consolidating following model support into causal_lm cpp demo:

• Qwen: Add Qwen CPP Pipeline #43
• ChatGLM: Wenyi5608 chatglm #48

[sammysun0711]

@Wovchena , @ilya-lavrenov, could you please review it?

[ilya-lavrenov]

as far as I remember, chatglm has different index for batch dimension
https://github.com/openvinotoolkit/openvino.genai/pull/48/files#diff-bceea0d0f5f31cbb280e385123d098ecddc2ee4bcc57a87b6789bce72a8b481cR85

[sammysun0711]

Yes, this PR is only for llama, will add chatglm support in another PR.

[Wovchena]

Apparently exporting a model as stateful removes any difference between the dimensions. I was able to run chatglm2-6b and chatglm3-6b using Wovchena#8 and get meaningful output. Here's the PR updating convert.py: #52. My next step is to rewrite this causal_lm as stateful as well, so we could decide if we need stateless approach at all.

[ilya-lavrenov]

I agree.
But we should wait until stateful models support is merged to HF, right?
It would be great to avoid when we refer to different commits to convert different models

[sammysun0711]

@Wovchena pls consider we may still need stateless models until both CPU and GPU plugin support stateful model.

[slyalin]

We are going to have stateful support for both CPU and GPU.

[slyalin]

We are considering to have separate samples for stateless models if they demonstrate popular architectures and require architecture-dependent kv-cache processing.

[slyalin]

But the main focus is on stateful models.

[ilya-lavrenov]

so, maybe we can merge chatglm sample as is and treat it as "sample for stateless model for popular architecture" ?
While current one will show stateful models and hence it become more generic

[Wovchena]

Greedy search with stateful model: sammysun0711#1 That eliminates a need to have special implementation for chatglm.

I also had that Idea to have stateless model with architecture, which doesn't fit into the general scenario. And there's still Qwen which may require that #43. I haven't tried it yet

[sammysun0711]

@Wovchena, @ilya-lavrenov , @slyalin thanks for your feedback. I've learn a lot from it.
Based on discussion above, I think I will change this PR scope with only following changes:

• Fix directory name typo
• Upgrade OpenVINO 23.2 in https://github.com/openvinotoolkit/openvino.genai/blob/master/text_generation/casual_lm/cpp/set_up_and_run.sh for CI

As @ilya-lavrenov suggested, we can keep causal_lm sample for generic stateful model enabling.
Meanwhile, as @slyalin suggested, we can add ChatGLM3 and Qwen as stateless model sample that demonstrate popular architectures and require architecture-dependent kv-cache processing.

• Qwen: Add Qwen CPP Pipeline #43 (Will update it with OV tokenizer)
• ChatGLM: Wenyi5608 chatglm #48

Please feel free to share any comments and feedback, thanks!