diff --git a/conversion/base.py b/conversion/base.py index 7cfbff3..e7a850a 100644 --- a/conversion/base.py +++ b/conversion/base.py @@ -2471,6 +2471,10 @@ class TextModel(ModelBase): # ref: https://huggingface.co/thinkingmachines/Inkling-2-Small res = "inkling2" + if chkhsh == "9b33e5d83df24061dcdb1f30e51583be2f0879fb36c2e13d3b03c037742561f3": + # ref: https://huggingface.co/Reson8/Rumble-2-Flash (pruned Qwen3-ASR tokenizer) + res = "qwen2" + if res is None: logger.warning("\n") logger.warning("**************************************************************************************") diff --git a/conversion/qwen3vl.py b/conversion/qwen3vl.py index 9458366..2693b5b 100644 --- a/conversion/qwen3vl.py +++ b/conversion/qwen3vl.py @@ -1,6 +1,7 @@ from __future__ import annotations import json +import os from typing import Any, Callable, Iterable, TYPE_CHECKING @@ -234,6 +235,23 @@ class Qwen3ASRMmprojModel(Qwen3OmniMmprojModel): has_audio_encoder = True has_vision_encoder = False + def get_audio_config(self) -> dict[str, Any] | None: + # The Transformers-5 checkpoint keeps audio_config at the top level (no thinker_config). + return self.global_config.get("audio_config") or self.global_config["thinker_config"].get("audio_config") + + @classmethod + def filter_tensors(cls, item): + name, gen = item + if name.startswith("model."): + name = name.replace("model.", "", 1) + if name.startswith("multi_modal_projector.linear_"): + # Transformers-5 layout: projector is multi_modal_projector.linear_1/2; the GGUF map + # (and llama.cpp's qwen3a graph) expects audio_tower.proj1/2. + name = "audio_tower.proj" + name.rsplit("_", 1)[1] + if "audio_tower." not in name: + return None + return MmprojModel.filter_tensors((name, gen)) + @ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration") @ModelBase.example("zai-org/GLM-4.1V-9B-Thinking", "zai-org/GLM-4.5V") @@ -390,11 +408,24 @@ class Qwen3ASRTextModel(Qwen3VLTextModel): def set_gguf_parameters(self): super().set_gguf_parameters() self.gguf_writer.add_num_deepstack_layers(0) + # Rumble's config drops mrope_section; Qwen3-ASR-0.6B uses [24, 20, 20, 0] and the + # audio/text positions are identical in every section, so this is plain RoPE. + if "mrope_section" not in self.rope_parameters: + self.gguf_writer.add_rope_dimension_sections([24, 20, 20, 0]) def set_vocab(self): super().set_vocab() - # fix chat template, use correct chatml format - self.gguf_writer.add_chat_template("{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}") + # llama-server runs the model with a media marker inside the user text, and the vendor + # template is written for typed content plus {% generation %}. RUMBLE_CHAT_TEMPLATE points + # at a llama.cpp-compatible adaptation (see the model repo); fall back to plain chatml. + template = None + template_path = os.environ.get("RUMBLE_CHAT_TEMPLATE") + if template_path and os.path.isfile(template_path): + with open(template_path, "r", encoding="utf-8") as f: + template = f.read() + if template is None: + template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}" + self.gguf_writer.add_chat_template(template) # correct BOS/EOS tokens with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f: tokenizer_config = json.load(f)