Fix bug of deepseek-v3 under DP+EP mode with large batchsize/seqlen (#6449)

2025-07-10 16:19:56 +08:00
parent 5748241549
commit 4a0d19198b
2 changed files with 6 additions and 6 deletions
--- a/python/sglang/srt/layers/quantization/fp8_kernel.py
+++ b/python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -160,8 +160,8 @@ def _per_token_group_quant_fp8_colmajor(
    """
    # Map the program id to the row of X and Y it should compute.
    g_id = tl.program_id(0)
-    y_ptr += g_id * group_size
-    y_q_ptr += g_id * group_size
+    y_ptr += g_id.to(tl.int64) * group_size
+    y_q_ptr += g_id.to(tl.int64) * group_size

    # Convert g_id the flattened block coordinate to 2D so we can index
    # into the output y_scales matrix