From 889521009bf9fde96eb1625812185330a71c0d93 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Sun, 31 May 2026 04:45:58 +0000 Subject: [PATCH] =?UTF-8?q?re-enable=20inverse=20RoPE=20(confirmed=20neces?= =?UTF-8?q?sary=20=E2=80=94=20without=20it=20output=20is=20garbage)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- single_shot_inference.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/single_shot_inference.py b/single_shot_inference.py index 05d05505..6f244374 100644 --- a/single_shot_inference.py +++ b/single_shot_inference.py @@ -405,9 +405,7 @@ def forward_layer(X_l, w, li, cfg, rope_cos, rope_sin, attn_out = attn_out.permute(1, 0, 2) # (T, n_h, hd) # -- Inverse RoPE on attention output (paper §2.3.3) -- - # NOTE: disabling for debugging — check if this is causing issues - # attn_out = apply_inverse_rope(attn_out, positions_dev, rope_cos, rope_sin, hd, rd) - attn_out = attn_out # No inverse RoPE for now + attn_out = apply_inverse_rope(attn_out, positions_dev, rope_cos, rope_sin, hd, rd) # -- Output projection: wo_a (grouped BMM) + wo_b (NVFP4) -- # wo_a: grouped linear, (n_h, hd) → (n_groups, o_rank) via BMM