Fix minor docs issues and fix metric requests

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,12 +3,16 @@
 import argparse
 import copy
 import itertools
 
 import torch
+import triton
+import triton.language as tl
+from typing import Optional, List, Dict
+import pandas as pd
 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
 
@@ -28,6 +32,104 @@
 _enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
 
 
+def benchmark_nvfp4_gemm_enhanced(
+    m: int,
+    n: int,
+    k: int,
+    use_cutlass: bool = True,
+    use_triton: bool = True,
+    dtype: torch.dtype = torch.float16
+) -> List[Dict]:
+    results = []
+
+    device = torch.device('cuda')
+    compute_capability = torch.cuda.get_device_capability(device)
+    sm_version = compute_capability[0] * 10 + compute_capability[1]
+
+    print(f"Device SM version: {sm_version}")
+
+    a = torch.randn(m, k, device=device, dtype=dtype)
+    b = torch.randn(k, n, device=device, dtype=dtype)
+
+    if use_cutlass and sm_version == 100:
+        try:
+            from vllm.experimental.kernels import cutlass_fp8_gemm
+
+            for _ in range(10):
+                out_cutlass = cutlass_fp8_gemm(a, b)
+
+            torch.cuda.synchronize()
+
+            import time
+            num_iterations = 100
+            start = time.time()
+
+            for _ in range(num_iterations):
+                out_cutlass = cutlass_fp8_gemm(a, b)
+
+            torch.cuda.synchronize()
+            end = time.time()
+
+            elapsed_ms = (end - start) * 1000 / num_iterations
+            flops = 2 * m * n * k
+            tflops = flops / (elapsed_ms / 1000) / 1e12
+
+            results.append({
+                'implementation': 'CUTLASS',
+                'elapsed_ms': elapsed_ms,
+                'tflops': tflops,
+                'm': m,
+                'n': n,
+                'k': k,
+                'sm_version': sm_version
+            })
+
+        except ImportError:
+            print("CUTLASS implementation not available")
+
+    if use_triton:
+        triton_result = benchmark_triton_nvfp4_gemm(m, n, k, dtype)
+        triton_result.update({
+            'implementation': 'Triton',
+            'm': m,
+            'n': n,
+            'k': k,
+            'sm_version': sm_version
+        })
+        results.append(triton_result)
+
+    for _ in range(10):
+        out_ref = torch.matmul(a, b)
+
+    torch.cuda.synchronize()
+
+    import time
+    num_iterations = 100
+    start = time.time()
+
+    for _ in range(num_iterations):
+        out_ref = torch.matmul(a, b)
+
+    torch.cuda.synchronize()
+    end = time.time()
+
+    elapsed_ms = (end - start) * 1000 / num_iterations
+    flops = 2 * m * n * k
+    tflops = flops / (elapsed_ms / 1000) / 1e12
+
+    results.append({
+        'implementation': 'PyTorch (FP16)',
+        'elapsed_ms': elapsed_ms,
+        'tflops': tflops,
+        'm': m,
+        'n': n,
+        'k': k,
+        'sm_version': sm_version
+    })
+
+    return results
+
+
 def _quant_weight_nvfp4(b: torch.Tensor, device: str):
     # Compute global scale for weight
     b_amax = torch.abs(b).max().to(torch.float32)

docs/community/meetups.md

@@ -1,6 +1,6 @@
 # Meetups
 
-We host regular meetups in San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below:
+We host regular meetups in the San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below:
 
 - [NYC vLLM Meetup](https://lu.ma/c1rqyf1f), May 7th, 2025. [[Slides]](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing)
 - [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day), April 3rd 2025. [[Slides]](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
@@ -17,4 +17,4 @@ We host regular meetups in San Francisco Bay Area every 2 months. We will share
 - [The second vLLM meetup](https://lu.ma/ygxbpzhl), with IBM Research, January 31st 2024. [[Slides]](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing) [[Video (vLLM Update)]](https://youtu.be/Y0C-DUvEnZQ) [[Video (IBM Research & torch.compile)]](https://youtu.be/m0dMtFLI-dg)
 - [The first vLLM meetup](https://lu.ma/first-vllm-meetup), with a16z, October 5th 2023. [[Slides]](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing)
 
-We are always looking for speakers and sponsors at San Francisco Bay Area and potentially other locations. If you are interested in speaking or sponsoring, please contact us at [[email protected]](mailto:[email protected]).
+We are always looking for speakers and sponsors in the San Francisco Bay Area and potentially other locations. If you are interested in speaking or sponsoring, please contact us at [[email protected]](mailto:[email protected]).

docs/models/extensions/fastsafetensor.md

@@ -2,4 +2,4 @@ Loading Model weights with fastsafetensors
 ===================================================================
 
 Using fastsafetensors library enables loading model weights to GPU memory by leveraging GPU direct storage. See [their GitHub repository](https://github.com/foundation-model-stack/fastsafetensors) for more details.
-For enabling this feature, set the environment variable ``USE_FASTSAFETENSOR`` to ``true``
+To enable this feature, set the environment variable ``USE_FASTSAFETENSOR`` to ``true``

vllm/v1/engine/output_processor.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import time
 import asyncio
 from collections.abc import Iterable
 from dataclasses import dataclass
@@ -10,6 +11,7 @@
 
 from vllm.outputs import (CompletionOutput, PoolingOutput,
                           PoolingRequestOutput, RequestOutput)
+from vllm.sequence import RequestMetrics
 from vllm.sampling_params import RequestOutputKind
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 from vllm.transformers_utils.tokenizer_group import TokenizerGroup
@@ -410,6 +412,14 @@
             if request_output := req_state.make_request_output(
                     new_token_ids, pooling_output, finish_reason, stop_reason,
                     kv_transfer_params, num_cached_tokens):
+                request_output.metrics = RequestMetrics(
+                    arrival_time=req_state.stats.arrival_time,
+                    last_token_time=req_state.stats.last_token_ts,
+                    first_scheduled_time=req_state.stats.scheduled_ts,
+                    first_token_time=req_state.stats.first_token_ts,
+                    time_in_queue=req_state.stats.scheduled_ts - req_state.stats.arrival_time,
+                    finished_time=time.monotonic()
+                )
                 if req_state.queue is not None:
                     # AsyncLLM: put into queue for handling by generate().
                     req_state.queue.put(request_output)

vllm/v1/engine/processor.py

@@ -247,7 +247,7 @@ def process_inputs(
                              f"is out of range [0, {data_parallel_size}).")
 
         if arrival_time is None:
-            arrival_time = time.time()
+            arrival_time = time.monotonic()
 
         # Process inputs, which includes:
         # 1. Tokenize text prompt, with LoRA request if one exists.