modeling_llama.py

from packaging import version
import transformers
from transformers import LlamaForCausalLM as HFLlamaForCausalLM
import warnings

_TRANSFORMERS_VERSION = version.parse(transformers.__version__)

print(f"[llama-yarn] Detected transformers version: {_TRANSFORMERS_VERSION}")

if _TRANSFORMERS_VERSION >= version.parse("5.0.0"):
    _patch_version = _TRANSFORMERS_VERSION
    print(
        f"[llama-yarn] Using default transformers implementation, "
        f"since transformers version {_patch_version} >= 5.0.0"
    )
    LlamaForCausalLM = HFLlamaForCausalLM

else:
    _patch_version = version.parse("4.46.3")
    print(f"[llama-yarn] Using transformers<5 patch (target version {_patch_version})")
    from .llama_yarn_patch_4x import LlamaForCausalLMYarn4x as LlamaForCausalLM

    if _TRANSFORMERS_VERSION == _patch_version:
        print(
            f"[llama-yarn] Patch version matches transformers exactly "
            f"({_TRANSFORMERS_VERSION})"
        )
    else:
        warnings.warn(
            "[llama-yarn] Patch version mismatch:\n"
            f"  transformers installed: {_TRANSFORMERS_VERSION}\n"
            f"  patch built for:        {_patch_version}\n"
            "The model may still work but compatibility is not guaranteed.",
            RuntimeWarning,
        )