fix(driver): keep Qwen3.5/3.6 linear-attention decay params in f32 so exp() runs on Metal

[zyma98]

Running a bf16 Qwen3.6 checkpoint on Metal warned:

48 graph node(s) fell back to CPU (no Metal kernel): UNARYx48

One per linear-attention layer: the gated-delta-rule block computes exp(A_log) directly on the bf16 linear_attn.A_log weight, and ggml-metal has no bf16 unary kernel, so every exp node fell back to CPU and split the graph. A_log/dt_bias slipped past the load-time f32 promotion, whose suffix match only covers *norm.weight/*.bias.

Fix, scoped to PieArch::Qwen3_5:

• model.cpp: promote linear_attn.{A_log,dt_bias} to f32 at load time ([num_v_heads] floats per layer, ~9 KiB total on 27B), and exempt them from the f32→bf16 downcast.
• graph_qwen3_5.cpp: defensive f32 cast before exp(A_log) for checkpoints that bypass the load-time upcast.

Verified on Qwen3.6-27B (M4 Pro): the warning is gone and the whole graph schedules onto MTL0. Qwen3-0.6B smoke test unchanged.

Summary by CodeRabbit

• Bug Fixes
  • Improved Qwen3.5 and Qwen3.6 linear-attention support on Metal.
  • Prevented unnecessary CPU fallback during model execution.
  • Improved compatibility for models using BF16 linear-decay parameters.

[coderabbitai]

Walkthrough

Qwen3.5 linear-attention decay weights are now retained as F32 during loading, and non-F32 lin_A_log tensors are cast to F32 before exponentiation.

Changes

Qwen3.5 precision handling

Layer / File(s)	Summary
Linear decay weight classification driver/portable/src/model.cpp	Identifies Qwen3.5 linear-attention A_log and dt_bias weights as small F32 weights and applies the consolidated upcast/downcast logic.
F32 exponentiation input driver/portable/src/graph_qwen3_5.cpp	Casts non-F32 L.lin_A_log tensors to F32 before applying ggml_exp.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: keeping Qwen3.5/3.6 linear-attention decay parameters in f32 to avoid Metal exp() fallback.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/qwen3.5-metal

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

driver/portable/src/graph_qwen3_5.cpp (1)

65-67: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Minor: repeated "cast-to-f32-if-needed" pattern.

dt_bias and A_log now both use the identical 3-line ternary-cast idiom. Could extract a small to_f32(ctx, t) helper to avoid duplicating this if a third such param appears later.

♻️ Optional helper extraction

+static ggml_tensor* to_f32(ggml_context* ctx, ggml_tensor* t) {
+    return (t->type == GGML_TYPE_F32) ? t : ggml_cast(ctx, t, GGML_TYPE_F32);
+}
+
 ...
-    auto* dt_bias = (L.lin_dt_bias->type == GGML_TYPE_F32)
-        ? L.lin_dt_bias
-        : ggml_cast(ctx, L.lin_dt_bias, GGML_TYPE_F32);
+    auto* dt_bias = to_f32(ctx, L.lin_dt_bias);
 ...
-    auto* A_log = (L.lin_A_log->type == GGML_TYPE_F32)
-        ? L.lin_A_log
-        : ggml_cast(ctx, L.lin_A_log, GGML_TYPE_F32);
+    auto* A_log = to_f32(ctx, L.lin_A_log);

Also applies to: 70-75

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@driver/portable/src/graph_qwen3_5.cpp` around lines 65 - 67, Extract the
repeated cast-to-F32 ternary used for dt_bias and A_log into a small to_f32(ctx,
tensor) helper, then replace both inline expressions with calls to that helper
while preserving the existing behavior for tensors already of type
GGML_TYPE_F32.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@driver/portable/src/graph_qwen3_5.cpp`:
- Around line 65-67: Extract the repeated cast-to-F32 ternary used for dt_bias
and A_log into a small to_f32(ctx, tensor) helper, then replace both inline
expressions with calls to that helper while preserving the existing behavior for
tensors already of type GGML_TYPE_F32.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: b5ebf8c4-597c-447e-a874-5729fc285b82

📥 Commits

Reviewing files that changed from the base of the PR and between 25ecc03 and dec555a.

📒 Files selected for processing (2)

• driver/portable/src/graph_qwen3_5.cpp
• driver/portable/src/model.cpp