fix: Support backwards-compatibility for "invocation_string" in adapter_config.json

While this has been replaced in the PEFT PR in favor of
alora_invocation_tokens, the existing adapters in the ibm-granite org on HF
use "invocation_string," so this will enable backwards compatibility and
enable testing now (before PEFT PR changes have percolated everywhere).

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <[email protected]>

Author: gabe-l-hart

convert_lora_to_gguf.py

@@ -12,7 +12,7 @@
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
-from transformers import AutoConfig
+from transformers import AutoConfig, AutoTokenizer
 
 import torch
 
@@ -373,7 +373,22 @@ def set_type(self):
             def set_gguf_parameters(self):
                 logger.debug("GGUF KV: %s = %d", gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
-                if alora_invocation_tokens := lparams.get("alora_invocation_tokens"):
+                alora_invocation_tokens = lparams.get("alora_invocation_tokens")
+                invocation_string = lparams.get("invocation_string")
+                if invocation_string and not alora_invocation_tokens:
+                    logger.debug("Tokenizing invocation_string -> alora_invocation_tokens")
+                    base_model_path_or_id = hparams.get("_name_or_path")
+                    try:
+                        tokenizer = AutoTokenizer.from_pretrained(base_model_path_or_id)
+                    except ValueError:
+                        logger.error("Unable to load tokenizer from %s", base_model_path_or_id)
+                        raise
+                    # NOTE: There's an off-by-one with the older aLoRAs where
+                    # the invocation string includes the "<|start_of_turn|>"
+                    # token, but the adapters themselves were trained to
+                    # activate _after_ that first token, so we drop it here.
+                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
+                if alora_invocation_tokens:
                     logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
                     self.gguf_writer.add_key_value(
                         gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS,