"""Host-side weight packing for attn_block_kernel (pure NumPy; no torch or NKI import).""" import numpy as np P = 128 D = 64 HALF = 32 QUAD = P // HALF def lane_perm(NH): """perm[g*128 + l] = feature index (within one projection) for lane group g = j*2 + par.""" NQ = NH // QUAD perm = np.zeros(NH * D, np.int64) for j in range(NQ): for par in range(2): g = j * 2 + par for bq in range(QUAD): h = j * QUAD + bq for i in range(HALF): perm[g * P + bq * HALF + i] = h * D + par * HALF + i return perm def pack(wqkv, bqkv, wo, bo, g, b, cos, sin, H, scale): """wqkv [3H, H], bqkv [3H] (qkv_proj, rows q|k|v), wo [H, H], bo [H] (o_proj), g/b [H] (attn_ln), cos/sin [L, D] (RoPE table rows for positions 0..L-1), scale = 1/sqrt(head_dim), folded into Wq/bq. Returns a dict of kernel inputs as fp32 NumPy arrays (cast wqk / wv / wo to the model dtype).""" NH = H // D perm = lane_perm(NH) wq, wk, wv = wqkv[:H], wqkv[H:2 * H], wqkv[2 * H:] bq, bk, bv = bqkv[:H], bqkv[H:2 * H], bqkv[2 * H:] wq_p, bq_p = wq[perm] * scale, bq[perm] * scale wk_p, bk_p = wk[perm], bk[perm] wqk = np.concatenate([wq_p, wk_p], 0).T.copy() # [H, 2H] bqk = np.concatenate([bq_p, bk_p]).reshape(-1, P).T.copy() # [128, 2H/128] lane_freq = np.arange(P) % HALF # lane l -> freq index cs = np.stack([cos[:, lane_freq].T, sin[:, lane_freq].T]).astype(np.float32) # [2, 128, L] bc = lambda v: np.tile(v.astype(np.float32), (P, 1)) return dict(wqk=wqk, bqk=bqk.astype(np.float32), wv=wv.T.copy(), bv=bc(bv), wo=wo.T.copy(), bo=bc(bo), g=bc(g), b=bc(b), cs=cs)