[DeepSeek-V4] Fuse UE8M0 scale rounding into FP8 group quantization (#26766)

Co-authored-by: liqichao <liqichao@baidu.com>
Co-authored-by: yhyang201 <yhyang201@gmail.com>

test/registered/jit/test_per_token_group_quant_8bit_v2.py

@@ -10,6 +10,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     fp8_dtype,
     fp8_max,
     fp8_min,
+    sglang_per_token_group_quant_fp8,
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
@@ -82,6 +83,32 @@ def test_v2_jit_matches_aot(dtype, num_tokens, hidden, fuse_silu_and_mul, scale_
     assert torch.equal(x_s, s_ref), "scales differ"
 
 
+@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
+@pytest.mark.parametrize("num_tokens", [1, 33, 128])
+@pytest.mark.parametrize("hidden", [128, 512, 4096, 7168])
+def test_sglang_per_token_group_quant_fp8_row_major_ue8m0(dtype, num_tokens, hidden):
+    """Row-major scale_ue8m0=True quantizes WITH the rounded (power-of-2) scale.
+    Verify: (1) scales are exact powers of 2, (2) dequant ≈ original within FP8 tolerance.
+    """
+    torch.manual_seed(num_tokens * 1000 + hidden)
+    x = torch.randn(num_tokens, hidden, device="cuda", dtype=dtype)
+
+    x_q, x_s = sglang_per_token_group_quant_fp8(x, G, scale_ue8m0=True)
+    torch.cuda.synchronize()
+
+    # Scales must be exact powers of 2
+    log2_s = torch.log2(x_s.abs())
+    assert torch.equal(log2_s, log2_s.round()), "scales are not power-of-2"
+
+    # Dequant should approximate original within FP8 precision
+    x_deq = x_q.float().view(num_tokens, -1, G) * x_s.unsqueeze(-1)
+    x_deq = x_deq.view(num_tokens, hidden)
+    rel_err = (x.float() - x_deq).abs() / (x.float().abs() + 1e-6)
+    assert (
+        rel_err.mean() < 0.05
+    ), f"mean relative dequant error too large: {rel_err.mean():.4f}"
+
+
 # Masked (EP-MoE) path: the v2 op only has a masked scheduler for the
 # column-major + ue8m0 + fused-silu+mul + masked combination. Input is 3D
 # [num_experts, tokens_padded, hidden*2]; only tokens < masked_m[e] are processed

test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py

@@ -41,7 +41,7 @@ class TestDSV4FlashFP4B200(
     """LowLatency recipe: TP=4, FP4 (mxfp4), EAGLE spec decoding."""
 
     gsm8k_accuracy_thres = 0.93
-    accept_length_thres = 2.6
+    accept_length_thres = 2.8
     bs_1_speed_thres = 220
 
     @classmethod

test/registered/models_e2e/test_deepseek_v4_flash_fp4_megamoe_b200.py

@@ -95,7 +95,7 @@ class TestDSV4FlashFP4B200W4A4MegaMoE(
     """Balanced recipe: TP=4, DP=4, MegaMoE."""
 
     gsm8k_accuracy_thres = 0.93
-    accept_length_thres = 2.6
+    accept_length_thres = 2.8
     bs_1_speed_thres = 100
 
     @classmethod