[MUSA] Resolve output garbage in Context Parallel on MusaFlashAttentionBackend (#23270)

Co-authored-by: zhiguo.qin <zhiguo.qin@mthreads.com>
2026-04-22 20:22:20 -07:00
parent 6490afe36e
commit 887d380ace
6 changed files with 128 additions and 61 deletions
@@ -112,15 +112,15 @@ srt_hpu = ["sglang[runtime_common]"]

 # https://docs.sglang.io/platforms/mthreads_gpu.md
 srt_musa = [
-    "sglang[runtime_common]",
-    "torch",
-    "torch_musa",
-    "torchada>=0.1.48",
-    "mthreads-ml-py",
-    "mate",
-    "mate-deep_gemm",
-    "mate-flash-attention",
-    "numpy<2.0",
+  "sglang[runtime_common]",
+  "torch",
+  "torch_musa",
+  "torchada>=0.1.50",
+  "mthreads-ml-py",
+  "mate>=0.2.0",
+  "deep-gemm>=0.1.3",
+  "flash_attn_3>=0.1.4",
+  "numpy<2.0",
 ]

 diffusion_musa = [