GPT-2 doesn't always have an lm_head

[steventrouble]

The GPT2 ggml models don't always have an lm_head, but llm assumes it does:

llm/crates/models/gpt2/src/lib.rs

Line 62 in 9a22269

let lm_head = tl.load("model/lm_head")?;

See ggml example repo where they mention lm_head is optional: https://github.com/ggerganov/ggml/blob/ee1b3727e60403012dd2b57d35b60558f4db66d8/examples/gpt-2/main.cpp#L364

[steventrouble]

I tried replacing it with

let lm_head = tl.load("model/lm_head").or_else(|_| tl.load("model/wte"))?;

But then I ran into a more ggml errors and had to move on

ggml_new_tensor_impl: not enough space in the context's memory pool (needed 252199424, available 250733056)

[LLukas22]

Yeah i also noticed that, we probably have to share the wte tensor as the context size is calculated via the combined size of all tensors.