PMAT-3477: Qwen3.5 (Gated DeltaNet hybrid) runs on CUDA — nine device kernels, Qwen35CudaModel, refusal sites lifted, apr run/chat/qa route to the GPU (#3090) by noahgift · Pull Request #3517 · paiml/aprender

PMAT-3477: Gated DeltaNet device kernels — conv1d+SiLU, per-head L2 norm, gates, delta-rule recurrence, gated RMSNorm, sigmoid gate — each against a host port of the CPU reference (#3090)

noahgift

and

claude

committed

5ebc79d

PMAT-3477: head_dim-256 decode attention, partial NEOX RoPE and q|gate split kernels for the Qwen3.5 attention layers (#3090)

noahgift

and

claude

committed

7591ff1

PMAT-3477: Qwen35CudaModel — executor wrappers for the six GDN kernels and the DeltaNet layer forward on the GPU, every DeltaNet layer matching the CPU forward on the real 0.8B file (#3090)

noahgift

and

claude

committed

58870fd

PMAT-3477: Qwen3.5 attention layers on the GPU and the full per-token forward — argmax matches the CPU at every position on the real 0.8B file (#3090)

noahgift

and

claude

committed

6474c45

PMAT-3477: qwen35 GPU refusal sites lifted in lockstep — capability gate, parity, ptx-map, quant whitelist and the SSM predicate now agree with the hybrid forward (#3090)

noahgift

and

claude

committed

d9459c7

PMAT-3477: Qwen35CudaModel pins the float GEMV (DP4A is catastrophic through the recurrence); parity restated to per-op exact ≤1e-3 plus e2e argmax equality with a measured cosine floor (#3090)

noahgift

and

claude

committed

1ed0adf

PMAT-3477: apr run/chat/qa route qwen35 to the GPU forward; the F2 guard judges the hybrid against its own CPU forward; no silent CPU fallback (#3090)

noahgift

and

claude

committed

b2bd0c5

PMAT-3477: Qwen35CudaModel per-token loop is sync-free — no per-layer stream syncs, LM head on device, one logits download; floor test no longer re-pins (#3090 review)

noahgift

and

claude

committed

9ce3dee

PMAT-3477: Gated DeltaNet with num_v_heads > num_k_heads (GQA in the recurrence) on the CPU forward — Qwen3.5 4B/9B/27B load and run (#3346, #3510)

noahgift

and

claude

committed

1a97abb

Merge remote-tracking branch 'origin/main' into PMAT-3477-gdn-gpu

noahgift

committed

48dffa9

PMAT-3477: claim-literal rows keep their lines — kernel_type.rs rebuilt from main with the nine GDN variants appended, speedup.rs leading blank restored; review quorum artifact for #3517

noahgift

and

claude

committed

b030e00

Revert "PMAT-3477: Gated DeltaNet with num_v_heads > num_k_heads (GQA in the recurrence) on the CPU forward — Qwen3.5 4B/9B/27B load and run (#3346, #3510)"

noahgift

committed

0f5aaab

Navigation Menu

Repository navigation

PMAT-3477: Qwen3.5 (Gated DeltaNet hybrid) runs on CUDA — nine device kernels, Qwen35CudaModel, refusal sites lifted, apr run/chat/qa route to the GPU (#3090) - #3517

PMAT-3477: Qwen3.5 (Gated DeltaNet hybrid) runs on CUDA — nine device kernels, Qwen35CudaModel, refusal sites lifted, apr run/chat/qa route to the GPU (#3090)#3517
noahgift merged 12 commits into
mainfrom
PMAT-3477-gdn-gpu

Commits on Sep 19, 2026

PMAT-3477: Gated DeltaNet device kernels — conv1d+SiLU, per-head L2 norm, gates, delta-rule recurrence, gated RMSNorm, sigmoid gate — each against a host port of the CPU reference (#3090)

PMAT-3477: head_dim-256 decode attention, partial NEOX RoPE and q|gate split kernels for the Qwen3.5 attention layers (#3090)

PMAT-3477: Qwen35CudaModel — executor wrappers for the six GDN kernels and the DeltaNet layer forward on the GPU, every DeltaNet layer matching the CPU forward on the real 0.8B file (#3090)

PMAT-3477: Qwen3.5 attention layers on the GPU and the full per-token forward — argmax matches the CPU at every position on the real 0.8B file (#3090)

PMAT-3477: qwen35 GPU refusal sites lifted in lockstep — capability gate, parity, ptx-map, quant whitelist and the SSM predicate now agree with the hybrid forward (#3090)

PMAT-3477: Qwen35CudaModel pins the float GEMV (DP4A is catastrophic through the recurrence); parity restated to per-op exact ≤1e-3 plus e2e argmax equality with a measured cosine floor (#3090)

PMAT-3477: apr run/chat/qa route qwen35 to the GPU forward; the F2 guard judges the hybrid against its own CPU forward; no silent CPU fallback (#3090)

PMAT-3477: Qwen35CudaModel per-token loop is sync-free — no per-layer stream syncs, LM head on device, one logits download; floor test no longer re-pins (#3090 review)

PMAT-3477: Gated DeltaNet with num_v_heads > num_k_heads (GQA in the recurrence) on the CPU forward — Qwen3.5 4B/9B/27B load and run (#3346, #3510)

Merge remote-tracking branch 'origin/main' into PMAT-3477-gdn-gpu

PMAT-3477: claim-literal rows keep their lines — kernel_type.rs rebuilt from main with the nine GDN variants appended, speedup.rs leading blank restored; review quorum artifact for #3517

Revert "PMAT-3477: Gated DeltaNet with num_v_heads > num_k_heads (GQA in the recurrence) on the CPU forward — Qwen3.5 4B/9B/27B load and run (#3346, #3510)"