feat: add Qwen3.5 support

[JohannesHa]

Bump vLLM to >=0.16.1.dev (nightly) which includes Qwen3.5 model support. This requires torch 2.10 (resolved from the existing >=2.9.0 pin), an updated flash-attn wheel built against torch 2.10, and version overrides for nvidia-cutlass-dsl and quack-kernels.

Bump transformers pin to 5c1c72b which includes a rope validation fix for Qwen3.5 (huggingface/transformers#44272).

Add a trainer monkey-patch for a transformers bug where Qwen3.5 passes 3D MRoPE position_ids to decoder layers instead of 2D text_position_ids, which breaks flash attention and causes NaN gradients. The upstream fix is pending: huggingface/transformers#44399

---

Note

^{Cursor Bugbot is generating a summary for commit 2767dea. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Autofix Details

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Duplicate nvidia-cutlass-dsl override renders new entry redundant
  • Removed the duplicate nvidia-cutlass-dsl>=4.4.1 constraint, keeping only the >=4.4.0.dev1 entry which allows both dev versions and stable releases.

Or push these changes by commenting:

@cursor push caadd1da72

Preview (caadd1da72)

diff --git a/pyproject.toml b/pyproject.toml
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -79,7 +79,6 @@
 # See: https://github.com/pytorch/pytorch/issues/166122
 override-dependencies = [
     "nvidia-cudnn-cu12>=9.15",
-    "nvidia-cutlass-dsl>=4.4.1",
     "transformers>=5.1.0.dev0",
     "nvidia-cutlass-dsl>=4.4.0.dev1",
     "quack-kernels>=0.2.7",

_{This Bugbot Autofix run was free. To enable autofix for future PRs, go to the Cursor dashboard.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Perf counter dense MLP uses wrong intermediate size for MoE

Low Severity

The fallback getattr(config, "intermediate_size", getattr(config, "moe_intermediate_size", 0)) sets intermediate_size to moe_intermediate_size for models that lack intermediate_size (like pure MoE models). This value is then used for dense_mlp_params on line 134. For MoE models with some dense layers (e.g., models with first_k_dense_replace), the dense layer intermediate size may differ from moe_intermediate_size, yielding an incorrect FLOP estimate.