I think one of the following should hold: 1. FA2 avoids physically reading padded/unwritten slots in partial KV blocks; or 2. FA2 is allowed to read them, but the invariant is documented and tested: those positions must be masked out before the loaded K/V values can affect output.

vllm - 💡(How to fix) Fix [Bug]: FA2 partial-block clamp can load unwritten padded KV slots [1 participants]

vllm2026-05-13 07:33:10

ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

vllm-project/vllm#42496•Fetched 2026-05-14 03:29:37

View on GitHub

Comments

Participants

Timeline

Reactions

Author

jongukc

Participants

jongukc

Timeline (top)

labeled ×1subscribed ×1

I observed a possible robustness or documentation issue in the FlashAttention-2 paged KV partial block path.

When the last KV block in an FA2 tile is only partially populated, resolve_thread_kv_page_slice_offset clamps extra threads to a fixed in-block row. This avoids reading past the block table, but it can also make those clamped threads physically load KV slots that were not written for the current request.

Current outputs appear to remain correct because those padded positions are outside the true sequence length and are masked out before they affect softmax. I found nearby comments documenting the clamp and the last-block mask separately, but I could not find a comment documenting the combined contract: clamped FA2 paged-KV threads may physically load padded/unwritten KV slots, and correctness relies on the seqlen_k-derived mask.

Root Cause

This appears operationally safe today because the attention mask removes those padded positions. The concern is mainly future maintenance: the correctness story depends on the mask being wired to the exact valid sequence length for every relevant FA2/paged-KV launch.

A future backend, FA kernel variant, or mask-boundary regression could make these physical reads observable. A related precedent is PR #30887, where an attention backend failed to fully mask out-of-window V values and those values could affect output.

Fix Action

Fix / Workaround

============================== CPU Info

Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 43 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 256 On-line CPU(s) list: 0-255 Vendor ID: AuthenticAMD Model name: AMD EPYC 7742 64-Core Processor CPU family: 23 Model: 49 Thread(s) per core: 2 Core(s) per socket: 64 Socket(s): 2 Stepping: 0 Frequency boost: enabled CPU(s) scaling MHz: 130% CPU max MHz: 2250.0000 CPU min MHz: 1500.0000 BogoMIPS: 4491.80 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es Virtualization: AMD-V L1d cache: 4 MiB (128 instances) L1i cache: 4 MiB (128 instances) L2 cache: 64 MiB (128 instances) L3 cache: 512 MiB (32 instances) NUMA node(s): 8 NUMA node0 CPU(s): 0-15,128-143 NUMA node1 CPU(s): 16-31,144-159 NUMA node2 CPU(s): 32-47,160-175 NUMA node3 CPU(s): 48-63,176-191 NUMA node4 CPU(s): 64-79,192-207 NUMA node5 CPU(s): 80-95,208-223 NUMA node6 CPU(s): 96-111,224-239 NUMA node7 CPU(s): 112-127,240-255 Vulnerability Gather data sampling: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection Vulnerability Spec rstack overflow: Mitigation; safe RET Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

Suggested fixes / mitigations

Code Example

Collecting environment information...
uv is set
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : version 4.3.2
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu129
Is debug build               : False
CUDA used to build PyTorch   : 12.9
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-5.15.0-1070-nvidia-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : 
GPU 0: NVIDIA A100-SXM4-40GB
  MIG 3g.20gb     Device  0:

Nvidia driver version        : 550.127.08
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        43 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               256
On-line CPU(s) list:                  0-255
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7742 64-Core Processor
CPU family:                           23
Model:                                49
Thread(s) per core:                   2
Core(s) per socket:                   64
Socket(s):                            2
Stepping:                             0
Frequency boost:                      enabled
CPU(s) scaling MHz:                   130%
CPU max MHz:                          2250.0000
CPU min MHz:                          1500.0000
BogoMIPS:                             4491.80
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es
Virtualization:                       AMD-V
L1d cache:                            4 MiB (128 instances)
L1i cache:                            4 MiB (128 instances)
L2 cache:                             64 MiB (128 instances)
L3 cache:                             512 MiB (32 instances)
NUMA node(s):                         8
NUMA node0 CPU(s):                    0-15,128-143
NUMA node1 CPU(s):                    16-31,144-159
NUMA node2 CPU(s):                    32-47,160-175
NUMA node3 CPU(s):                    48-63,176-191
NUMA node4 CPU(s):                    64-79,192-207
NUMA node5 CPU(s):                    80-95,208-223
NUMA node6 CPU(s):                    96-111,224-239
NUMA node7 CPU(s):                    112-127,240-255
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec rstack overflow:   Mitigation; safe RET
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.6
[pip3] mypy-extensions==1.1.0
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.9.1.4
[pip3] nvidia-cuda-cupti-cu12==12.9.79
[pip3] nvidia-cuda-nvcc-cu12==12.9.86
[pip3] nvidia-cuda-nvrtc-cu12==12.9.86
[pip3] nvidia-cuda-runtime-cu12==12.9.79
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.4.1.4
[pip3] nvidia-cufile-cu12==1.14.1.1
[pip3] nvidia-curand-cu12==10.3.10.19
[pip3] nvidia-cusolver-cu12==11.7.5.82
[pip3] nvidia-cusparse-cu12==12.5.10.65
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.2
[pip3] nvidia-cutlass-dsl-libs-base==4.4.2
[pip3] nvidia-ml-py==13.595.45
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.9.86
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.9.79
[pip3] open-clip-torch==2.32.0
[pip3] pytorch-lightning==2.6.1
[pip3] pyzmq==27.1.0
[pip3] segmentation-models-pytorch==0.5.0
[pip3] sentence-transformers==5.4.1
[pip3] terratorch==1.2.7
[pip3] torch==2.10.0+cu129
[pip3] torch-c-dlpack-ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchgeo==0.9.0
[pip3] torchmetrics==1.9.0
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==5.5.3
[pip3] transformers-stream-generator==0.0.5
[pip3] triton==3.6.0
[pip3] tritonclient==2.68.0
[pip3] vector-quantize-pytorch==1.28.2
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.19.2.dev17+gddd036cbd (git sha: ddd036cbd)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  	[4mGPU0	NIC0	NIC1	NIC2	NIC3	NIC4	NIC5	NIC6	NIC7	NIC8	NIC9	CPU Affinity	NUMA Affinity	GPU NUMA ID[0m
GPU0	 X 	SYS	SYS	SYS	SYS	SYS	SYS	PXB	PXB	SYS	SYS	80-95,208-223	5		N/A
NIC0	SYS	 X 	PXB	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC1	SYS	PXB	 X 	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC2	SYS	SYS	SYS	 X 	PXB	SYS	SYS	SYS	SYS	SYS	SYS				
NIC3	SYS	SYS	SYS	PXB	 X 	SYS	SYS	SYS	SYS	SYS	SYS				
NIC4	SYS	SYS	SYS	SYS	SYS	 X 	PXB	SYS	SYS	SYS	SYS				
NIC5	SYS	SYS	SYS	SYS	SYS	PXB	 X 	SYS	SYS	SYS	SYS				
NIC6	PXB	SYS	SYS	SYS	SYS	SYS	SYS	 X 	PXB	SYS	SYS				
NIC7	PXB	SYS	SYS	SYS	SYS	SYS	SYS	PXB	 X 	SYS	SYS				
NIC8	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	 X 	PIX				
NIC9	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	PIX	 X 				

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7
  NIC8: mlx5_8
  NIC9: mlx5_9

==============================
     Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=/var/run/nvidia-container-devices
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root

---

int64_t block_row_offset = tidx / kGmemThreadsPerRow * kGmemRowsPerThread;

if (partial_block_size) {
    auto final_row_offset = std::max(*partial_block_size - 1, 0);
    auto final_thread_row_offset =
      ceil_div(final_row_offset, kGmemRowsPerThread) * kGmemRowsPerThread;
    block_row_offset = std::min(
        block_row_offset, int64_t(final_thread_row_offset));
}

const int64_t global_row_offset = block_row_offset + n_block * kBlockN;
const int64_t page_offset = global_row_offset % page_block_size;
const int64_t virtual_page_idx = global_row_offset / page_block_size;
const int phys_block = block_table[virtual_page_idx];

RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary>

Collecting environment information...
uv is set
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.4 LTS (x86_64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0
Clang version                : Could not collect
CMake version                : version 4.3.2
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu129
Is debug build               : False
CUDA used to build PyTorch   : 12.9
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0] (64-bit runtime)
Python platform              : Linux-5.15.0-1070-nvidia-x86_64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : 
GPU models and configuration : 
GPU 0: NVIDIA A100-SXM4-40GB
  MIG 3g.20gb     Device  0:

Nvidia driver version        : 550.127.08
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        43 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               256
On-line CPU(s) list:                  0-255
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7742 64-Core Processor
CPU family:                           23
Model:                                49
Thread(s) per core:                   2
Core(s) per socket:                   64
Socket(s):                            2
Stepping:                             0
Frequency boost:                      enabled
CPU(s) scaling MHz:                   130%
CPU max MHz:                          2250.0000
CPU min MHz:                          1500.0000
BogoMIPS:                             4491.80
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es
Virtualization:                       AMD-V
L1d cache:                            4 MiB (128 instances)
L1i cache:                            4 MiB (128 instances)
L2 cache:                             64 MiB (128 instances)
L3 cache:                             512 MiB (32 instances)
NUMA node(s):                         8
NUMA node0 CPU(s):                    0-15,128-143
NUMA node1 CPU(s):                    16-31,144-159
NUMA node2 CPU(s):                    32-47,160-175
NUMA node3 CPU(s):                    48-63,176-191
NUMA node4 CPU(s):                    64-79,192-207
NUMA node5 CPU(s):                    80-95,208-223
NUMA node6 CPU(s):                    96-111,224-239
NUMA node7 CPU(s):                    112-127,240-255
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec rstack overflow:   Mitigation; safe RET
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.6
[pip3] mypy-extensions==1.1.0
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.9.1.4
[pip3] nvidia-cuda-cupti-cu12==12.9.79
[pip3] nvidia-cuda-nvcc-cu12==12.9.86
[pip3] nvidia-cuda-nvrtc-cu12==12.9.86
[pip3] nvidia-cuda-runtime-cu12==12.9.79
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.4.1.4
[pip3] nvidia-cufile-cu12==1.14.1.1
[pip3] nvidia-curand-cu12==10.3.10.19
[pip3] nvidia-cusolver-cu12==11.7.5.82
[pip3] nvidia-cusparse-cu12==12.5.10.65
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.2
[pip3] nvidia-cutlass-dsl-libs-base==4.4.2
[pip3] nvidia-ml-py==13.595.45
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.9.86
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.9.79
[pip3] open-clip-torch==2.32.0
[pip3] pytorch-lightning==2.6.1
[pip3] pyzmq==27.1.0
[pip3] segmentation-models-pytorch==0.5.0
[pip3] sentence-transformers==5.4.1
[pip3] terratorch==1.2.7
[pip3] torch==2.10.0+cu129
[pip3] torch-c-dlpack-ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchgeo==0.9.0
[pip3] torchmetrics==1.9.0
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==5.5.3
[pip3] transformers-stream-generator==0.0.5
[pip3] triton==3.6.0
[pip3] tritonclient==2.68.0
[pip3] vector-quantize-pytorch==1.28.2
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.19.2.dev17+gddd036cbd (git sha: ddd036cbd)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  	[4mGPU0	NIC0	NIC1	NIC2	NIC3	NIC4	NIC5	NIC6	NIC7	NIC8	NIC9	CPU Affinity	NUMA Affinity	GPU NUMA ID[0m
GPU0	 X 	SYS	SYS	SYS	SYS	SYS	SYS	PXB	PXB	SYS	SYS	80-95,208-223	5		N/A
NIC0	SYS	 X 	PXB	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC1	SYS	PXB	 X 	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC2	SYS	SYS	SYS	 X 	PXB	SYS	SYS	SYS	SYS	SYS	SYS				
NIC3	SYS	SYS	SYS	PXB	 X 	SYS	SYS	SYS	SYS	SYS	SYS				
NIC4	SYS	SYS	SYS	SYS	SYS	 X 	PXB	SYS	SYS	SYS	SYS				
NIC5	SYS	SYS	SYS	SYS	SYS	PXB	 X 	SYS	SYS	SYS	SYS				
NIC6	PXB	SYS	SYS	SYS	SYS	SYS	SYS	 X 	PXB	SYS	SYS				
NIC7	PXB	SYS	SYS	SYS	SYS	SYS	SYS	PXB	 X 	SYS	SYS				
NIC8	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	 X 	PIX				
NIC9	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	SYS	PIX	 X 				

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7
  NIC8: mlx5_8
  NIC9: mlx5_9

==============================
     Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=/var/run/nvidia-container-devices
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root

</details>

🐛 Describe the bug

Summary

I observed a possible robustness or documentation issue in the FlashAttention-2 paged KV partial block path.

Question for maintainers

Is this an expected FA2 implementation detail?

In other words, are clamped threads allowed to physically read padded/unwritten slots as long as the seqlen_k-derived mask guarantees those positions cannot affect output?

If this is intended, I think a short comment near the clamp would make the contract clearer for future FA2 / FA3 / paged-KV backend work.

Actual behavior

The relevant helper is in flash-attention/csrc/flash_attn/src/utils.h:

int64_t block_row_offset = tidx / kGmemThreadsPerRow * kGmemRowsPerThread;

if (partial_block_size) {
    auto final_row_offset = std::max(*partial_block_size - 1, 0);
    auto final_thread_row_offset =
      ceil_div(final_row_offset, kGmemRowsPerThread) * kGmemRowsPerThread;
    block_row_offset = std::min(
        block_row_offset, int64_t(final_thread_row_offset));
}

const int64_t global_row_offset = block_row_offset + n_block * kBlockN;
const int64_t page_offset = global_row_offset % page_block_size;
const int64_t virtual_page_idx = global_row_offset / page_block_size;
const int phys_block = block_table[virtual_page_idx];

The clamp protects the block-table index, but the clamped page_offset can correspond to a padded slot inside the final physical block.

The model output was still correct in my tests, which is consistent with FA2's per-position causal/padding mask assigning zero probability to those padded positions.

Related comments I found

I may be missing something, but the comments I found seem to document the pieces separately rather than the full contract:

flash-attention/csrc/flash_attn/src/utils.h says the partial-block clamp adjusts the row offset to avoid reading past the end of the block table.
flash-attention/csrc/flash_attn/src/flash_fwd_kernel.h says the last K/V block needs score masking when K/V length is not a multiple of kBlockN, and that some smem tiles do not need clearing because the scores will be masked.
flash-attention/csrc/flash_attn/src/mask.h applies -INFINITY for positions outside max_seqlen_k.

Those comments explain why the current implementation appears correct, but I could not find one that explicitly says the clamped paged-KV load may hit padded/unwritten slots inside the final physical block.

I also saw the FlexAttention documentation/comment about garbage values in unused block-table positions. I think this is a separate case: the FA2 partial-block issue is about a valid final physical block whose padded in-block slots may not have been written.

Expected behavior

I think one of the following should hold:

FA2 avoids physically reading padded/unwritten slots in partial KV blocks; or
FA2 is allowed to read them, but the invariant is documented and tested: those positions must be masked out before the loaded K/V values can affect output.

Why this matters

Suggested fixes / mitigations

Possible options:

Add a comment near the FA2 partial-block clamp explaining that clamped threads may read padded slots and that correctness relies on the per-position mask.
Add a regression test that poisons padded/unwritten KV slots in a partial block and verifies outputs are unchanged.
If the extra safety is worth the cost, route clamped padded-position reads to a known-safe sentinel/null block or otherwise avoid loading unwritten slots.

Notes

I am not claiming that current FA2 outputs are wrong. In my tests, outputs remained correct.

This report is about (1) whether the partial-block clamp's masked physical reads are an intended contract, and (2) whether that contract should be documented or tested.

Before submitting a new issue...

Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

Vote matrix · Quick signals

Works

Did the solution work? Tap to confirm.

Easy Fix

Was it a quick fix?

Time Saver

Did it save you time?

Blocking

Was it severely blocking?

Common Issue

Are others likely hitting this too?

Flaky / Intermittent

Is it intermittent?

Verified / Reproducible

Can you reproduce it reliably?

FAQ

Expected behavior

I think one of the following should hold:

FA2 avoids physically reading padded/unwritten slots in partial KV blocks; or
FA2 is allowed to read them, but the invariant is documented and tested: those positions must be masked out before the loaded K/V values can affect output.

#api #tensor shape #autograd error #model save/load #environment variable

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

vllm - 💡(How to fix) Fix [Bug]: FA2 partial-block clamp can load unwritten padded KV slots [1 participants]

Recommended Tools

GitHub issue graph ai analysis

Root Cause

Fix Action

Fix / Workaround

============================== CPU Info

Suggested fixes / mitigations

Code Example

Your current environment

🐛 Describe the bug

Summary

Question for maintainers

Actual behavior

Related comments I found

Expected behavior

Why this matters

Suggested fixes / mitigations

Notes

Before submitting a new issue...

FAQ

Expected behavior

Still need to ship something?

TRENDING

vllm - 💡(How to fix) Fix [Bug]: FA2 partial-block clamp can load unwritten padded KV slots [1 participants]

Recommended Tools

GitHub issue graph ai analysis

Root Cause

Fix Action

Fix / Workaround

============================== CPU Info

Suggested fixes / mitigations

Code Example

Your current environment

🐛 Describe the bug

Summary

Question for maintainers

Actual behavior

Related comments I found

Expected behavior

Why this matters

Suggested fixes / mitigations

Notes

Before submitting a new issue...

FAQ

Expected behavior

Still need to ship something?

RELATED_DISCOVERY

TRENDING