[Model] Add Grok-2 (#31847)

Signed-off-by: dangoldbj <dangoldbj23@gmail.com>
This commit is contained in:
Bijaya Dangol
2026-01-08 13:59:48 +01:00
committed by GitHub
parent 18d4e481d0
commit 59d260f5e4
8 changed files with 777 additions and 20 deletions

View File

@@ -31,6 +31,7 @@ logger = init_logger(__name__)
_VLLM_TOKENIZERS = {
"deepseek_v32": ("deepseek_v32", "DeepseekV32Tokenizer"),
"grok2": ("grok2", "Grok2Tokenizer"),
"hf": ("hf", "CachedHfTokenizer"),
"mistral": ("mistral", "MistralTokenizer"),
}
@@ -151,6 +152,17 @@ def resolve_tokenizer_args(
if len(files_list) > 0:
tokenizer_mode = "mistral"
# Try to use Grok2 tiktoken tokenizer if possible
if tokenizer_mode == "auto":
allow_patterns = ["tokenizer.tok.json"]
files_list = list_filtered_repo_files(
model_name_or_path=str(tokenizer_name),
allow_patterns=allow_patterns,
revision=revision,
)
if len(files_list) > 0:
tokenizer_mode = "grok2"
# Fallback to HF tokenizer
if tokenizer_mode == "auto":
tokenizer_mode = "hf"