From 8cf4d73c846f8c38237679d756de519eb64c7e67 Mon Sep 17 00:00:00 2001 From: ashwini rathi Date: Sat, 10 Oct 2026 06:36:24 +0530 Subject: [PATCH] [XPU] Fix XPU CI breaks from #41105 (KDA chunk_offsets) and #42588 (int4 test tp_group) (#43334) --- .../xpu/kernels/fla/chunk_delta_h.py | 11 ++++++----- test/registered/xpu/test_xpu_int4_dense.py | 12 ++++++------ 2 files changed, 12 insertions(+), 11 deletions(-) diff --git a/python/sglang/srt/hardware_backend/xpu/kernels/fla/chunk_delta_h.py b/python/sglang/srt/hardware_backend/xpu/kernels/fla/chunk_delta_h.py index 0102ab4cad..f478d3ebaf 100644 --- a/python/sglang/srt/hardware_backend/xpu/kernels/fla/chunk_delta_h.py +++ b/python/sglang/srt/hardware_backend/xpu/kernels/fla/chunk_delta_h.py @@ -271,6 +271,7 @@ def chunk_gated_delta_rule_fwd_h( inplace_update: bool = True, track_state: Optional[torch.Tensor] = None, track_chunk_idx: Optional[torch.Tensor] = None, + chunk_offsets: Optional[torch.Tensor] = None, ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: if not inplace_update: raise NotImplementedError( @@ -298,11 +299,11 @@ def chunk_gated_delta_rule_fwd_h( if cu_seqlens is None: N, NT, chunk_offsets = B, triton.cdiv(T, BT), None else: - N, NT, chunk_offsets = ( - len(cu_seqlens) - 1, - len(chunk_indices), - prepare_chunk_offsets(cu_seqlens, BT), - ) + # A caller-built chunk_offsets (CUDA-graph capture) must not go through + # prepare_chunk_offsets: its tensor_cache would replay capture-time values. + N, NT = len(cu_seqlens) - 1, len(chunk_indices) + if chunk_offsets is None: + chunk_offsets = prepare_chunk_offsets(cu_seqlens, BT) assert K <= 256, "current kernel does not support head dimension larger than 256." h = k.new_empty(B, NT, H, V, K) diff --git a/test/registered/xpu/test_xpu_int4_dense.py b/test/registered/xpu/test_xpu_int4_dense.py index 2ea8b665e3..96c2893cae 100644 --- a/test/registered/xpu/test_xpu_int4_dense.py +++ b/test/registered/xpu/test_xpu_int4_dense.py @@ -1,10 +1,10 @@ """Numeric unit tests for the XPU int4 *dense* linear kernels (GPTQ / AWQ).""" import unittest +from types import SimpleNamespace import torch -from sglang.srt.runtime_context import get_parallel from sglang.srt.utils import is_xpu from sglang.test.ci.ci_register import register_xpu_ci from sglang.test.test_utils import CustomTestCase @@ -196,8 +196,8 @@ class TestXPUInt4DenseKernel(CustomTestCase): layer.g_idx = torch.nn.Parameter(g_idx, requires_grad=False) kernel = GPTQXPULinearKernel(_gptq_config(gs, desc_act, fmt)) - with get_parallel().override(tp_size=tp_size): - kernel.process_weights_after_loading(layer) + layer.tp_group = SimpleNamespace(world_size=tp_size) + kernel.process_weights_after_loading(layer) out = kernel.apply(layer, x) self.assertEqual(tuple(out.shape), (m, n)) @@ -239,9 +239,9 @@ class TestXPUInt4DenseKernel(CustomTestCase): (2, r"tp_size=2.*--tp-size 1"), ): with self.subTest(tp_size=tp_size): - with get_parallel().override(tp_size=tp_size): - with self.assertRaisesRegex(NotImplementedError, pattern): - kernel.process_weights_after_loading(layer) + layer.tp_group = SimpleNamespace(world_size=tp_size) + with self.assertRaisesRegex(NotImplementedError, pattern): + kernel.process_weights_after_loading(layer) def test_gptq_group_aligned_shard_allows_tensor_parallel(self): # Whole-group shards stay representable, so TP is only rejected when a