[eduardoworrel]

Problem

Qwen3_5ForCausalLM throws in generate() for a text-only qwen3_5 checkpoint (model_type: "qwen3_5_text", no vision_config) when the ONNX graph exposes a position_ids input:

Cannot read properties of undefined (reading 'spatial_merge_size')
    at Function.get_rope_index (modeling_qwen2_vl.js)
    at Function.prepare_inputs_for_generation

Repro: AutoModelForCausalLM.from_pretrained(<qwen3_5_text checkpoint whose decoder exports position_ids>) then model.generate({ ...tokenizer('Hello'), max_new_tokens: 4 }). No image involved.

Cause

prepare_inputs_for_generation gates on the export:

const session = this.sessions["decoder_model_merged"] ?? this.sessions["model"];
if (!session.inputNames.includes("position_ids")) {
    return model_inputs;                       // checkpoint without the input: never reaches the crash
}
[model_inputs.position_ids, ...] = this.get_rope_index(...);   // checkpoint with it: crashes

get_rope_index (inherited from the Qwen2-VL lineage) then does:

const { vision_config } = this.config;
const spatial_merge_size = vision_config.spatial_merge_size ?? 2;   // dereferenced unconditionally
if (image_grid_thw || video_grid_thw) {                             // the value is only needed here

The ?? 2 fallback cannot apply — reading .spatial_merge_size on undefined throws first. Qwen3_5ForCausalLM extends Qwen3_5ForConditionalGeneration, which extends Qwen3VLForConditionalGeneration, so a text-only checkpoint inherits a multimodal get_rope_index that assumes a vision tower exists.

Checkpoints whose export omits position_ids are unaffected, which is why this only surfaces with certain exports.

Fix

-        const spatial_merge_size = vision_config.spatial_merge_size ?? 2;
+        const spatial_merge_size = vision_config?.spatial_merge_size ?? 2;

No behaviour change when vision_config exists; the value is only consumed inside the if (image_grid_thw || video_grid_thw) branch.

Verification

Same checkpoint, same call:

• Without the patch: generate() throws Cannot read properties of undefined (reading 'spatial_merge_size') from get_rope_index.
• With the patch and a rebuild: the call executes get_rope_index without throwing and proceeds past that point.

Scope of this patch is only the undefined dereference.

References

• In transformers (Python) Qwen3_5ForCausalLM subclasses Qwen3_5PreTrainedModel and composes Qwen3_5TextModel — a sibling of the multimodal class, whose get_rope_index handles the text-only case. The JS port inherits the multimodal one instead.
• Qwen2VLForConditionalGeneration documents that "CausalLM variants (e.g., Qwen2VLForCausalLM) extend this class but load only embed_tokens + decoder_model_merged (no vision_encoder)" — this dereference defeats that for a checkpoint with no vision_config at all.

[eduardoworrel]

Closing: the dereference is only reachable when the ONNX graph exports a position_ids input, and optimum-onnx (the supported exporter) does not export one for this architecture. Revisiting if that changes.