feat(aprender-serve): forward_qwen3_moe_cuda_traced — M-MOE-SUB-2 step (b)

[noahgift]

Summary

GPU traced sibling of forward_qwen3_moe_cuda. Mirrors the CPU traced sibling (PR #1516 forward_qwen3_moe_traced_with_plan) but routes per-layer MoE FFN through the GPU dispatch so apr trace --gpu --json --payload --save-tensor can run the same SaveTensorPlan against both CPU and GPU forward paths, capture per-stage activations at MoeRouter + MoeFfnOut, and bisect the M-GPU-MOE-1.4 NaN/Inf poisoning to its first divergence point.

Why

§40 + qwen3-moe-forward-gpu-v1 v1.4.0 amendment_history: heavy qwen3_moe_gpu_parity test on RTX 4090 against cached 17.3 GB Qwen3-Coder GGUF produces ALL 151936 logits as NaN at lm_head. CPU forward on the same input produces finite output. Steps 1-9 of the GPU forward are CPU-identical to the CPU forward (precondition checks, embedding, attention norm, QKV, RoPE, attention, output projection, FFN norm). Step 10 is the only GPU-specific stage — the MoE FFN. With this PR + #1522 (helper) + the existing CPU traced sibling, M-MOE-SUB-3 can apr trace --gpu the same plan against both forward bodies and pinpoint where the GPU NaN enters.

What ships

• OwnedQuantizedModelCuda::forward_qwen3_moe_cuda_traced(token_ids, moe_layers, num_experts, num_experts_per_tok, moe_intermediate, data) -> Result<ForwardTrace> — no-plan delegate.
• forward_qwen3_moe_cuda_traced_with_plan(..., plan: Option<&SaveTensorPlan>) -> Result<ForwardTrace> — plan-aware body.
• New file crates/aprender-serve/src/gguf/cuda/forward_qwen3_moe_cuda_traced.rs (~430 LOC including doc-comments).
• include!() registered in cuda/uses.rs.
• Lib-only signature drift gate test.

Hot path safety

Production forward_qwen3_moe_cuda is unchanged byte-for-byte. The per-token loop dispatches the GPU MoE FFN identically to production for non-capture positions; the LAST sequence position uses moe_ffn_forward_layer_cuda_with_router (PR #1522) when the plan selects MoeRouter or MoeFfnOut so router weights can be emitted without recomputation.

M-MOE-SUB-2 status after this PR

Sub-step	What	Status
(a) CPU body	forward_qwen3_moe_traced_with_plan	SHIPPED PR #1516
(a) CLI wireup	apr trace --save-tensor for GGUF MoE	SHIPPED PR #1521
(b) GPU body	forward_qwen3_moe_cuda_traced[_with_plan]	THIS PR
(c) CPU helper	moe_ffn_forward_layer_with_router	SHIPPED PR #1507
(c.gpu) GPU helper	moe_ffn_forward_layer_cuda_with_router	SHIPPED PR #1522

M-MOE-SUB-2 will be COMPLETE when this lands. M-MOE-SUB-3 (heavy parity test) is next.

Falsifier

FALSIFY-MOE-SUB-002 (byte-identity preservation for existing stages). This PR does not discharge it (heavy parity test required at M-MOE-SUB-3); it provides the GPU traced sibling that will run the test.

Test plan

• cargo test -p aprender-serve --features cuda --lib forward_qwen3_moe_cuda_traced_signature → 1 passed (signature drift gate)
• cargo fmt --check (file-level)
• cargo check -p aprender-serve --features cuda → finishes
• Auto-merge once required checks pass
• M-MOE-SUB-3 heavy parity test will exercise byte-identity vs production on lambda-vector RTX 4090

🤖 Generated with Claude Code