[Feature] Batch Invariant: Support DeepGEMM and Blackwell (#27127)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

tests/v1/generation/test_batch_invariance.py

@@ -10,9 +10,9 @@ import torch
 from vllm import LLM, SamplingParams
 from vllm.platforms import current_platform
 
-hopper_only = pytest.mark.skipif(
-    not (current_platform.is_cuda() and current_platform.is_device_capability(90)),
-    reason="Requires CUDA and Hopper (SM90)",
+skip_unsupported = pytest.mark.skipif(
+    not (current_platform.is_cuda() and current_platform.has_device_capability(90)),
+    reason="Requires CUDA and >= Hopper (SM90)",
 )
 
 
@@ -74,7 +74,7 @@ def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str:
     return base_prompt
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.timeout(1000)
 def test_v1_generation_is_deterministic_across_batch_sizes_with_needle():
     """
@@ -219,7 +219,7 @@ def _extract_step_logprobs(request_output):
     return None, None
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER"])
 @pytest.mark.forked
 def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(backend):
@@ -434,7 +434,7 @@ def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(backend):
         pytest.fail(msg)
 
 
-@hopper_only
+@skip_unsupported
 def test_simple_generation():
     """
     Simple test that runs the model with a basic prompt and prints the output.
@@ -480,7 +480,7 @@ def test_simple_generation():
             llm.shutdown()
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER"])
 @pytest.mark.forked
 def test_logprobs_WITHOUT_batch_invariance_should_FAIL(backend):
@@ -707,7 +707,7 @@ def test_logprobs_WITHOUT_batch_invariance_should_FAIL(backend):
             os.environ["VLLM_BATCH_INVARIANT"] = old_value
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("backend", ["FLASH_ATTN"])
 @pytest.mark.forked
 def test_decode_logprobs_match_prefill_logprobs(backend):

tests/v1/generation/test_rms_norm_batch_invariant.py

@@ -14,13 +14,13 @@ from vllm.model_executor.layers.batch_invariant import rms_norm as triton_rms_no
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.platforms import current_platform
 
-hopper_only = pytest.mark.skipif(
-    not (current_platform.is_cuda() and current_platform.is_device_capability(90)),
-    reason="Requires CUDA and Hopper (SM90)",
+skip_unsupported = pytest.mark.skipif(
+    not (current_platform.is_cuda() and current_platform.has_device_capability(90)),
+    reason="Requires CUDA and >= Hopper (SM90)",
 )
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("batch_size", [1, 4, 16, 64])
 @pytest.mark.parametrize("hidden_size", [512, 2048, 4096, 8192])
 @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
@@ -69,7 +69,7 @@ def test_rms_norm_batch_invariant_vs_standard(
     )
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("batch_size", [1, 16, 128])
 @pytest.mark.parametrize("seq_len", [1, 32, 512])
 @pytest.mark.parametrize("hidden_size", [2048, 4096])
@@ -111,7 +111,7 @@ def test_rms_norm_3d_input(batch_size: int, seq_len: int, hidden_size: int):
     )
 
 
-@hopper_only
+@skip_unsupported
 def test_rms_norm_numerical_stability():
     """
     Test RMS norm numerical stability with extreme values.
@@ -171,7 +171,7 @@ def test_rms_norm_numerical_stability():
         )
 
 
-@hopper_only
+@skip_unsupported
 def test_rms_norm_formula():
     """
     Test that RMS norm follows the correct mathematical formula.
@@ -204,7 +204,7 @@ def test_rms_norm_formula():
     )
 
 
-@hopper_only
+@skip_unsupported
 @pytest.mark.parametrize("hidden_size", [128, 1024, 4096, 16384])
 def test_rms_norm_different_hidden_sizes(hidden_size: int):
     """
@@ -242,7 +242,7 @@ def test_rms_norm_different_hidden_sizes(hidden_size: int):
     )
 
 
-@hopper_only
+@skip_unsupported
 def test_rms_norm_determinism():
     """
     Test that batch-invariant RMS norm produces deterministic results.

vllm/model_executor/layers/quantization/fp8.py

@@ -41,6 +41,7 @@ from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
 )
+from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
 from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
 from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
     FlashinferMoeBackend,
@@ -94,9 +95,11 @@ from vllm.platforms import current_platform
 from vllm.scalar_type import scalar_types
 from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (
+    fp8_gemm_nt,
     get_col_major_tma_aligned_tensor,
     is_deep_gemm_e8m0_used,
     is_deep_gemm_supported,
+    should_use_deepgemm_for_fp8_linear,
 )
 from vllm.utils.flashinfer import has_flashinfer_moe
 
@@ -539,8 +542,34 @@ class Fp8LinearMethod(LinearMethodBase):
         x: torch.Tensor,
         bias: torch.Tensor | None = None,
     ) -> torch.Tensor:
-        # If batch invariant mode is enabled, dequantize and use BF16 compute
+        # if batch invariant mode is enabled, prefer DeepGEMM FP8 path
+        # we will use BF16 dequant when DeepGEMM is not supported.
         if vllm_is_batch_invariant():
+            if self.block_quant and should_use_deepgemm_for_fp8_linear(
+                torch.bfloat16, layer.weight, None
+            ):
+                # use group quant consistent with block size across K
+                assert self.act_q_group_shape is not None
+                q_input, input_scale = QuantFP8(
+                    False,
+                    self.act_q_group_shape,
+                    column_major_scales=True,
+                )(x)
+
+                output_2d = torch.empty(
+                    (q_input.shape[0], layer.weight.shape[0]),
+                    dtype=torch.bfloat16,
+                    device=q_input.device,
+                )
+                fp8_gemm_nt(
+                    (q_input, input_scale),
+                    (layer.weight, layer.weight_scale),
+                    output_2d,
+                )
+                if bias is not None:
+                    output_2d = output_2d + bias
+                return output_2d
+
             # Dequantize FP8 weights to BF16
             weight_fp8 = layer.weight.to(torch.bfloat16)
             weight_scale = layer.weight_scale.to(torch.bfloat16)
@@ -555,9 +584,30 @@ class Fp8LinearMethod(LinearMethodBase):
 
                 N, K = weight_fp8.shape
 
-                # Scale is stored transposed: [num_blocks_k, num_blocks_n]
-                # We need to transpose it to [num_blocks_n, num_blocks_k] first
-                weight_scale = weight_scale.t()
+                # determine expected number of blocks along N and K
+                num_blocks_n = (N + block_n - 1) // block_n
+                num_blocks_k = (K + block_k - 1) // block_k
+
+                # scale layout may be [num_blocks_n, num_blocks_k]
+                # or [num_blocks_k, num_blocks_n] depending on backend
+                if weight_scale.dim() != 2:
+                    raise RuntimeError(
+                        f"FP8 block scale must be 2D, got {tuple(weight_scale.shape)}"
+                    )
+
+                scale_rows, scale_cols = weight_scale.shape
+                if (scale_rows, scale_cols) == (num_blocks_k, num_blocks_n):
+                    if num_blocks_n == num_blocks_k:
+                        # ambiguous square case, warn and skip transpose
+                        logger.warning(
+                            "Batch-invariant FP8: square block-scale %dx%d; "
+                            "skipping transpose to avoid misorientation.",
+                            scale_rows,
+                            scale_cols,
+                        )
+                    else:
+                        # clear KN -> transpose to NK
+                        weight_scale = weight_scale.t()
 
                 # Expand scale to match weight dimensions
                 # scale_expanded should have shape [N, K]