From 5e781900fb512a18861121517a1cdcee5a221bc8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=87=91=E9=BB=84=E8=89=B2=E8=91=A1=E8=90=84=E7=90=83?=
 =?UTF-8?q?=E5=90=9B=E5=90=9B?= <yueyuan@amd.com>
Date: Tue, 3 Mar 2026 19:05:32 +0800
Subject: [PATCH] Revert "perf(ROCm): optimize chunked CE loss num_warps for
 RDNA GPUs (#4123)" (#4139)

This reverts commit 721bf4852a802da44b78e3b000edd691e5af1d38.
---
 unsloth/kernels/cross_entropy_loss.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/unsloth/kernels/cross_entropy_loss.py b/unsloth/kernels/cross_entropy_loss.py
index 1bb1691776..d92229314f 100644
--- a/unsloth/kernels/cross_entropy_loss.py
+++ b/unsloth/kernels/cross_entropy_loss.py
@@ -22,7 +22,6 @@ from .utils import (
     triton_cast,
     torch_gpu_device,
     is_cdna,
-    is_rdna,
 )
 from transformers.models.llama.modeling_llama import logger
 from unsloth_zoo.utils import Version
@@ -365,7 +364,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
                     SOFTCAP = logit_softcapping,
                     DO_LOGIT_SCALING = DO_LOGIT_SCALING,
                     LOGIT_SCALE = logit_scaling,
-                    num_warps = 16 if is_cdna() or is_rdna() else 32,
+                    num_warps = 32 if not is_cdna() else 16,
                 )
             # logsumexp(chunked_logsumexp) - x
             # Do the -x separately