remove extra files

Author: rahul-tuli

.gitignore

@@ -146,6 +146,7 @@ venv.bak/
 
 # mkdocs documentation
 /site
+docs/argparse
 docs/examples
 
 # mypy
@@ -202,5 +203,3 @@ shellcheck*/
 
 # Ignore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*
-local/
-*.patch

benchmarks/benchmark_dataset.py

@@ -324,6 +324,9 @@ def sample(
         input_low = int(real_input_len * (1 - range_ratio))
         input_high = int(real_input_len * (1 + range_ratio))
         output_low = int(output_len * (1 - range_ratio))
+        # Ensure the lower bound for output length is at least 1 to prevent
+        # sampling 0 tokens, which can cause request failures.
+        output_low = max(output_low, 1)
         output_high = int(output_len * (1 + range_ratio))
 
         # Add logging for debugging

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -0,0 +1,141 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.scalar_type import scalar_types
+from vllm.triton_utils import triton
+
+if not current_platform.has_device_capability(100):
+    raise RuntimeError("NVFP4 requires compute capability of 10.0 (Blackwell)")
+
+
+FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
+FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "nvfp4": dict(no_a_quant=False, enabled=True),
+    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_nvfp4(b: torch.Tensor, device: str):
+    # Compute global scale for weight
+    b_amax = torch.abs(b).max().to(torch.float32)
+    b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
+    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    return b_fp4, scale_b_fp4, b_global_scale
+
+
+def build_nvfp4_runner(cfg, a, b, dtype, device):
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
+
+    # Compute global scale for activation
+    # NOTE: This is generally provided ahead-of-time by the model checkpoint.
+    a_amax = torch.abs(a).max().to(torch.float32)
+    a_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / a_amax
+
+    # Alpha for the GEMM operation
+    alpha = 1.0 / (a_global_scale * b_global_scale)
+
+    if cfg["no_a_quant"]:
+        # Pre-quantize activation
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+
+        def run():
+            return ops.cutlass_scaled_fp4_mm(
+                a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+            )
+
+        return run
+
+    # Quantize activation on-the-fly
+    def run():
+        a_fp4, scale_a_fp4 = ops.scaled_fp4_quant(a, a_global_scale)
+        return ops.cutlass_scaled_fp4_mm(
+            a_fp4, b_fp4, scale_a_fp4, scale_b_fp4, alpha, dtype
+        )
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs NVFP4 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_nvfp4_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -0,0 +1,98 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import itertools
+from typing import Callable
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
+from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.triton_utils import triton
+
+
+# TODO(luka): use standalone_compile utility
+def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
+    def inner(*args):
+        torch._dynamo.mark_dynamic(args[arg_index], dim_index)
+        return fn(*args)
+
+    return inner
+
+
+torch._dynamo.config.recompile_limit = 8888
+compilation_config = CompilationConfig(custom_ops=["none"])
+with set_current_vllm_config(VllmConfig(compilation_config=compilation_config)):
+    torch_per_token_quant_fp8 = torch.compile(
+        QuantFP8(False, GroupShape.PER_TOKEN),
+        fullgraph=True,
+        dynamic=False,  # recompile for different shapes
+    )
+
+    # First dim is explicitly dynamic to simulate vLLM usage
+    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
+
+
+def cuda_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input)
+
+
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between Triton and CUDA implementations."""
+    device = torch.device("cuda")
+    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
+
+    torch_out, torch_scale = torch_per_token_quant_fp8(x)
+    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
+
+    if torch.allclose(
+        cuda_out.to(torch.float32), torch_out.to(torch.float32), rtol=1e-3, atol=1e-5
+    ) and torch.allclose(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+
+
+batch_size_range = [1, 16, 32, 64, 128]
+seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+
+configs = list(itertools.product(batch_size_range, seq_len_range))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size", "seq_len"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["torch", "cuda"],
+        line_names=["Torch", "CUDA"],
+        styles=[("blue", "-"), ("green", "-")],
+        ylabel="us",
+        plot_name="per-token-dynamic-quant-fp8-performance",
+        args={},
+    )
+)
+def benchmark_quantization(batch_size, seq_len, provider):
+    dtype = torch.float16
+    device = torch.device("cuda")
+
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch":
+        fn = lambda: torch_per_token_quant_fp8(x.clone())
+    elif provider == "cuda":
+        fn = lambda: cuda_per_token_quant_fp8(x.clone())
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)
+
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_moe.py

@@ -86,6 +86,9 @@ def benchmark_config(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
         w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_deep_gemm:
+        # we use the default block shape for deepgemm
+        block_quant_shape = [128, 128]
     if use_fp8_w8a8:
         if block_quant_shape:
             block_n, block_k = block_quant_shape[0], block_quant_shape[1]