[gg-rebased] perf(nf3): integrate Grid188 hybrid decode

tests/quantization/test_nvfp4_nf3_hybrid.py

@@ -4,12 +4,15 @@
 import pytest
 import torch
 
+from vllm.config.quantization import resolve_quantization_config
 from vllm.model_executor.layers.quantization import get_quantization_config
 from vllm.model_executor.layers.quantization.nvfp4_nf3_hybrid import (
     NvFp4Nf3HybridConfig,
+    _combined_tier_local_descriptors,
     _read_hybrid_keys,
     _unpack_nf3_codes,
 )
+from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Dynamic
 
 
 @pytest.mark.parametrize(
@@ -65,6 +68,21 @@ def test_config_rejects_missing_hybrid_bit_map():
         )
 
 
+def test_config_accepts_dense_mxfp8_online_overlay():
+    resolved = resolve_quantization_config(
+        "nvfp4_nf3_hybrid",
+        {
+            "linear": {"weight": "mxfp8"},
+            "ignore": ["re:.*kv_b_proj"],
+        },
+    )
+
+    assert resolved is not None
+    assert resolved.linear is not None
+    assert resolved.linear.weight == kMxfp8Dynamic
+    assert resolved.ignore == ["re:.*kv_b_proj"]
+
+
 def test_unpack_nf3_codes():
     expected = torch.tensor([[[0, 1, 2, 3, 4, 5, 6, 7]]], dtype=torch.int32)
     word = sum(int(code) << (index * 3) for index, code in enumerate(expected[0, 0]))
@@ -74,3 +92,20 @@ def test_unpack_nf3_codes():
     )
 
     torch.testing.assert_close(_unpack_nf3_codes(packed, size_k=8), expected)
+
+
+def test_grid188_tier_descriptors_encode_exact_partition():
+    remap = {
+        **{global_id: (0, global_id) for global_id in range(64)},
+        **{global_id: (1, global_id - 64) for global_id in range(64, 256)},
+    }
+
+    descriptors = _combined_tier_local_descriptors(remap)
+
+    assert descriptors[:64] == list(range(64))
+    assert descriptors[64:] == [0x100 | local_id for local_id in range(192)]
+
+
+def test_grid188_tier_descriptors_reject_incomplete_partition():
+    with pytest.raises(ValueError, match="does not cover all 256"):
+        _combined_tier_local_descriptors({0: (0, 0)})

tests/v1/attention/test_indexer_dcp_localize.py

@@ -306,6 +306,19 @@ def test_get_dcp_local_seq_lens_preserves_mtp_bounds_shape():
     torch.testing.assert_close(actual, expected)
 
 
+def test_get_dcp_local_seq_lens_rank_does_not_allocate_cuda_scalar(monkeypatch):
+    seq_lens = torch.tensor([17, 33, 65], dtype=torch.int32)
+
+    def fail_tensor_allocation(*args, **kwargs):
+        raise AssertionError("rank localization must not allocate a scalar tensor")
+
+    monkeypatch.setattr(torch, "tensor", fail_tensor_allocation)
+
+    actual = get_dcp_local_seq_lens(seq_lens, dcp_size=4, dcp_rank=2)
+
+    assert actual.tolist() == [4, 8, 16]
+
+
 def test_get_dcp_local_seq_lens_must_run_after_decode_expansion():
     world = 2
     rank = 1

vllm/config/quantization.py

@@ -157,7 +157,14 @@ def _coerce_spec(cls, v: Any, info: ValidationInfo) -> Any:
 # which the checkpoint explicitly leaves unquantized (shared experts via
 # `shared_experts`, other dense linears via `linear`).
 _MODELOPT_ONLINE_OVERLAY_NAMES = frozenset(
-    {"modelopt", "modelopt_fp4", "modelopt_mxfp8", "modelopt_mixed", "mxfp4"}
+    {
+        "modelopt",
+        "modelopt_fp4",
+        "modelopt_mxfp8",
+        "modelopt_mixed",
+        "mxfp4",
+        "nvfp4_nf3_hybrid",
+    }
 )
 
 

vllm/envs.py

@@ -63,6 +63,7 @@
     VLLM_USE_B12X_FP8_GEMM: bool = False
     VLLM_USE_B12X_WO_PROJECTION: bool = False
     VLLM_USE_B12X_MOE: bool = False
+    VLLM_NF3_GRID188_DECODE: bool = True
     VLLM_USE_B12X_MINIMAX_M3_MSA: bool = False
     VLLM_USE_B12X_DCP_A2A: bool = False
     VLLM_DCP_PROJECT_BEFORE_MERGE: bool = False
@@ -1091,6 +1092,10 @@ def _resolve_rust_frontend_path() -> str | None:
     # Use b12x for FP4 MoE experts.
     # This is opt-in while the b12x subsystems are brought over one at a time.
     "VLLM_USE_B12X_MOE": lambda: bool(int(os.getenv("VLLM_USE_B12X_MOE", "0"))),
+    # Exact TP4 GLM-5.2 E64-NVFP4/E192-NF3 one-grid decode specialization.
+    "VLLM_NF3_GRID188_DECODE": lambda: bool(
+        int(os.getenv("VLLM_NF3_GRID188_DECODE", "1"))
+    ),
     # Use b12x for MiniMax M3's block-sparse MSA attention.
     # This is opt-in while page-128 MSA support is integrated.
     "VLLM_USE_B12X_MINIMAX_M3_MSA": lambda: bool(