Ling-3.0-tiny-GLM-5.3-Flash-surgery

Ling → GLM-5.3-Flash surgery

Experimental 7.80B text checkpoint, saved with stock Transformers save_pretrained(). Weight and tokenizer donor: inclusionAI/Ling-3.0-tiny, transferred into the native GLM-5.3-Flash architecture through weight surgery. Full donor tokenizer and untied embedding/output tables retained.

This model is intended for testing, continued training, and post-training experiments. Generation quality is unstable. Known limitations include:

• Repetition loops: generation can get stuck repeating text.
• Morphological errors: malformed word forms and incorrect inflections.
• Unintended language switching / language mixing: Chinese characters can appear unexpectedly in otherwise English output.
• Arithmetic errors, including an observed 6 × 7 → 49 response.

from transformers import AutoTokenizer, Glm5NextForConditionalGeneration

repo = "imvladikon/Ling-3.0-tiny-GLM-5.3-Flash-surgery"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = Glm5NextForConditionalGeneration.from_pretrained(repo)

Tested with Transformers 5.16.1, including ten English generations using these plain loading calls. No custom model code or auto_map. Only text weights are supplied. The stock loader reports missing model.visual.* weights for a tiny unused vision scaffold (depth 0, width 32). Image/video inference is unsupported. Context configured to 2048; sparse indexer selection remains untrained.

Checkpoint layout

Tensors carry the names released GLM-5.3-Flash uses: one tensor per expert (mlp.experts.7.gate_proj.weight), separate q_conv1d / k_conv1d / v_conv1d, and hc_attn_fn rather than attn_hc.fn — 9510 tensors. Transformers fuses them during loading through the glm5_next entry of its conversion_mapping, so from_pretrained needs nothing extra.

Those fused names are what the modules hold in memory, which means save_pretrained() writes the fused form back out (688 tensors). Transformers reads that form again, but nothing else does: Megatron-Bridge and the Megatron converters map from the released names, because the fused layout is an implementation detail that was never published. After fine-tuning and saving this model, rewrite the result before handing it to anything but Transformers:

python to_released_layout.py --source ./my-finetune --destination ./my-finetune-released

An earlier revision of this repository was published in the fused form. This one was rewritten and checked on Transformers 5.17.0: every one of the 688 fused tensors reassembles bitwise from the 9510 released ones, the key set matches a released GLM-5.3-Flash index exactly, and English, Russian and multi-turn prompts generate.

The following ten examples demonstrate basic English generation. Questions are abbreviated; answers are verbatim. Generation used do_sample=False, max_new_tokens=40, and enable_thinking=False in the chat template. These examples provide limited evidence of overall quality and generation stability.

Choosing the chat template: Ling (default) or Flash

The repository contains one parameterized chat_template.jinja. Transformers loads it automatically with AutoTokenizer.from_pretrained(repo); no custom Python helper or separate template download is needed.

Pass template_variant="ling" or template_variant="flash" directly to tokenizer.apply_chat_template. Omitting the argument selects Ling.

from transformers import AutoTokenizer

repo = "imvladikon/Ling-3.0-tiny-GLM-5.3-Flash-surgery"
tokenizer = AutoTokenizer.from_pretrained(repo)
messages = [{"role": "user", "content": "What is the capital of France?"}]

prompt = tokenizer.apply_chat_template(
    messages,
    template_variant="ling",  # default; change to "flash" for the Flash format
    enable_thinking=False,
    add_generation_prompt=True,
    tokenize=False,
)

enable_thinking=False ends the generation prompt with <think></think> in both formats; True (the default) leaves <think> open. This controls the prompt, not a guarantee of the model's generated behavior. Optional tools=tools passes function schemas. In conversation history, tool-call arguments must be dictionaries and responses from tools must be strings. Tool execution remains the caller's responsibility. reasoning_effort="low", "high", or "max" applies only to the Flash format.

For tokenized model inputs, use tokenize=True, return_dict=True, return_tensors="pt". If you tokenize the rendered string separately, use add_special_tokens=False.

Generation stop tokens

The repository's generation_config.json defaults to Ling: EOS is <|endoftext|> / <|role_end|>, and <|role_end|> is allowed to be generated. A Jinja argument does not change the generation configuration. When selecting Flash, also configure its native stop tokens and restore its suppression of <|role_end|>:

from transformers import GenerationConfig

flash_config = GenerationConfig.from_pretrained(repo)
flash_config.eos_token_id = tokenizer.convert_tokens_to_ids(
    ["<|endoftext|>", "<|user|>", "<|observation|>"]
)
for field in ("suppress_tokens", "begin_suppress_tokens"):
    values = getattr(flash_config, field, None)
    if values is not None:
        setattr(flash_config, field, [i for i in values if i not in flash_config.eos_token_id])
role_end_id = tokenizer.convert_tokens_to_ids("<|role_end|>")
flash_config.suppress_tokens = sorted(set(flash_config.suppress_tokens or []) | {role_end_id})

# Pair this config with inputs rendered using template_variant="flash":
# output = model.generate(**inputs, generation_config=flash_config, max_new_tokens=256)

These options switch prompt formatting and generation stops; they do not change model weights. The generation examples above were recorded before this template update. Template rendering, tokenization, tool-message formatting, and save/reload were checked separately without loading model weights.