Upload qvac-gpu.js with huggingface_hub

qvac-gpu.js

@@ -980,7 +980,10 @@ export async function createQvacGPU(manifest, fetchTensor, cap = 64, eos = 2, st
   const P_mmQ3B = bits === 3 && q3f ? pipe(mmQ3BigKernel(), "mmQ3B") : null;   // big q3f (lm_head)
   const P_mmT2B = hasT2 ? pipe(mmT2Kernel(false, 16, 16, 4), "mmT2B") : null, P_mmT2Badd = hasT2 ? pipe(mmT2Kernel(true, 16, 16, 4), "mmT2Badd") : null;
   const P_mmT2RB = hasT2R ? pipe(mmT2RKernel(false, 16, 16, 4), "mmT2RB") : null, P_mmT2RBadd = hasT2R ? pipe(mmT2RKernel(true, 16, 16, 4), "mmT2RBadd") : null;
-  if (subNorm && (stream || manifest.moe)) throw new Error("bitnet sub-norm path is resident-only for now");
+  // DRAFT: dense sub-norm (BitNet) STREAMING is enabled — the unfused layerBody applies attn/ffn sub-norm
+  // (rms over the resident Nrm weights) with ws(role) from R[role], and mmW dispatches t2 (Falcon-E proves it);
+  // combined with the t2 stream packing (t2stream), dense sub-norm streams. MoE sub-norm stays resident-only.
+  if (subNorm && manifest.moe) throw new Error("bitnet sub-norm MoE path is resident-only for now");
   // fused ternary layer (resident only): qkv 3→1 pass, gate/up 2→1, both ropes 1, act⊕sub-norm 1
   // measured policy (A/B walls, 5-run medians): fusion wins on the sub-norm (BitNet) family
   // (12.3 vs 13.2 ms/tok on bitnet-2b) and LOSES on llama-arch ternary (24.5 vs 19.5 on