[model, ops] feat: shared-prefix training for Qwen3.5

docs/examples/qwen3_5.md

@@ -253,6 +253,39 @@ bash train.sh tasks/train_vlm.py configs/multimodal/qwen3_5_moe/qwen3_5_moe_vl_m
 EP-local expert count is divisible by the EP-FSDP size; otherwise VeOmni logs
 a warning and falls back to the communication path.
 
+## Shared-prefix training
+
+In GRPO-style RL each prompt is sampled `n` times and the `n` rollouts are trained as `n` packed
+`prompt + response` sequences, so the prompt is computed `n` times. With shared-prefix training the
+text backbone computes each prompt once per micro-batch: GatedDeltaNet layers fork the recurrent
+state at the end of the prompt, softmax-attention layers reuse the prompt's keys and values, and the
+LM head scores each shared position once. Outputs and gradients match packed training to within the
+difference that re-batching the same sequences produces by itself.
+
+The work drops from `n * (P + R)` tokens to `P + n * R`, and on Qwen3.5-4B about 84% of that ratio is
+realised. Long prompts gain most; under roughly 1k prompt tokens there is little to save.
+
+It is a field on the **model config**, off by default, and it needs a non-eager linear-attention
+backend: `triton` on Ascend, `fla` on GPU.
+
+```json
+{ "text_config": { "shared_prefix_training": true } }
+```
+
+Set it in the `config.json` of the model directory the trainer loads. Nothing else changes: detection
+is internal to the patched modeling, so a trainer driving VeOmni needs no flag of its own and no
+adapter code. When nothing in a micro-batch shares a prefix, or the batch carries image or video
+inputs, the ordinary packed path runs unchanged.
+
+What the caller has to get right is that a micro-batch holds **whole rollout groups**, i.e. a
+multiple of `n` sequences, with no reordering between packing and the forward. A group split across
+micro-batches is still correct, only slower, since just the rollouts that land together share work.
+With `verl` (`model_engine=veomni`) that means the legacy trainer with group-aware batch balancing,
+`use_dynamic_bsz` off, and `ppo_micro_batch_size_per_gpu` a multiple of `rollout.n`; the config field
+travels through `actor_rollout_ref.model.hf_config_path`, which is why no verl-side change is needed.
+
+Current limits: text-only prompts; dense Qwen3.5; no Ulysses SP, MTP or KV cache (each raises).
+
 ## Ulysses Sequence Parallelism
 
 Qwen3.5 supports Ulysses sequence parallelism for both its softmax attention layers and