Total Reported
323
11 applications
73.1% confirmed or fixed
Confirmed (Awaiting Fix)
154
47.7%
Triaged by maintainers
Fixed Upstream
82
25.4%
Patched in releases
Under Review
54
16.7%
Awaiting maintainer triage
Rejected
33
10.2%
Closed as invalid or duplicate
Reported Bugs by Application
Confirmed (Awaiting Fix)
Fixed
Pending
Rejected
Resolution Status
73.1% Confirmed + Fixed236
Confirmed & Fixed
Awaiting Fix 154 (47.7%)
Fixed 82 (25.4%)
Pending 54 (16.7%)
Rejected 33 (10.2%)
Classification by Bug Type
No reported bugs matched your filter or search query.
PyTorch
Deep Learning
91 bugs
41 Confirmed (Awaiting Fix)
28 Fixed
2 Pending
20 Rejected
41 Confirmed (Awaiting Fix)
28 Fixed
2 Pending
20 Rejected
torch.logcumsumexp runtime error inconsistency on CPU and CUDA (also documentation missing constraint)
torch.nn.functional.conv_transpose2d produces inconsistent output on CPU and CUDA
`torch.nn.functional.conv_transpose2d` has inconsistent handling of `float16` overflow on CPU
Inconsistent float16 overflow behavior between CPU and CUDA devices
FPE when using `torch.lcm_` with int32 tensor and int16 scalar
Inconsistent overflow handling in `torch.clamp_min` between CPU and CUDA for float16 tensors
FPE Crash in torch.fmod with large negative integer modulo value
Inconsistent Error Handling Between CPU and CUDA for `cosine_similarity`
Inconsistent Error Handling in `torch.addcmul` with `int16` Tensors
`torch.native_channel_shuffle` crashes with Floating Point Exception when given large integer parameter
FPE when calling `torch.pixel_shuffle()` with empty tensors and large upscale_factor
Segmentation fault when converting sparse COO tensor with complex values to dense
`torch.combinations` exhibits excessive memory usage and hangs for moderate `n` and `r` due to `n^r`
torch.dequantize result inconsistent on CPU and GPU
torch.aminmax reducing over all dims should not accept keepdim
Memory Corruption in `torch.batch_norm_update_stats`
torch.nn.InstanceNorm3d producing inconsistent output for float16 tensors on CPU and GPU
torch.dot has inconsistent support for int64 (long)
torch.nn.InstanceNorm2d throws "mixed dtype" error with track_running_stats set to True
torch.nn.ReflectionPad2d throws CUDA error on large tensors
torch.logcumsumexp runtime error inconsistency on CPU and CUDA (also documentation missing constraint)
`resize_` updates tensor shape metadata even when storage resize fails, creating corrupted "Zombie" tensors
`torch.nn.functional.pixel_shuffle` throws SIGFPE (Floating Point Exception) with large upscale_factor
`torch.nn.NLLLoss` has inconsistent error handling: CUDA skips type check for empty inputs vs CPU
torch.pixel_shuffle CPU-GPU crash with massive upscale factor
`torch.fbgemm_linear_quantize_weight` crashes with segmentation fault (SIGSEGV) with a cuda tensor
`torch._stack` crashes with segmentation fault (`SIGSEGV`) on empty tensors
`torch.native_channel_shuffle` crashes with FPE on GPU when `groups=0`
`torch.arange` triggers `SIGFPE` (Floating Point Exception) instead of RuntimeError when writing float results to `int64` out tensor
`torch.special.zeta` returns incorrect output with `uint16` on CUDA, along with many other `torch.special` APIs
`torch.linalg.slogdet` does not propagate `NaN` in CUDA
torch.logdet produces incorrect results for singular matrices on CUDA vs CPU
Segmentation fault (core dumped) in `torch._remove_batch_dim`
Floating point exception on torch.lcm
`torch.floor_divide` crashes with Floating point exception (`SIGFPE`) when dividing `INT64_MIN`
Segmentation Fault Due to Empty Stride Argument in `torch._C._nn.slow_conv3d`
Feedback about torch.addcdiv
torch.nn.functional.softshrink throws overflow error on CUDA but not on CPU
`torch.nn.functional.rrelu` crashes on CPU with `training=True` when `lower` or `upper` is set to `inf`
`torch.native_channel_shuffle` throws FPE when `groups` > size of the second dimension of the input tensor
`torch.nn.LPPool2d` throws inconsistent error on CPU and GPU
torch.nn.functional.normalize producing nan values with a large p value and tensor of complex numbers
Inconsistent behavior and misleading error message for `torch.nanmean()` with complex dtypes
Inconsistent behavior between CPU and GPU implementations of `torch.arange`
`torch.batch_norm` shows inconsistent error behavior between CPU and GPU
IInconsistent Error Handling in `torch.fused_moving_avg_obs_fake_quant` Between CPU and GPU Implementations
`lp_pool1d` behavior inconsistency between CPU and GPU
Segmentation fault when calling `torch.choose_qparams_optimized()` with empty tensors and extreme num_bins value
FPE in `torch.remainder`
torch.acos produces different values using cpu and cuda for input with torch.complex128 dtype
torch.nn.ReplicationPad and torch.nn.ReflectionPad does not support float16
torch.sigmoid producing nan for tensor of negative complex numbers on cpu
torch.asin returns incorrect value with complex input on cpu
torch.histc not checking min > max on cuda for int8 tensors
`torch.nn.LogSoftmax` produces inf value on CPU but not on GPU
`torch.nn.functional.celu` produces inconsistent output on CPU vs CUDA
`torch.floor_divide` causes inconsistent precision
`torch.fmin` has inconsistent overflow behavior on CPU and GPU
`torch.det` returns inconsistent results for complex tensor on CPU vs GPU
Inconsistent behavior between CPU and GPU implementations of `torch.Tensor.put_` method
Inconsistent Complex torch.Tensor.asin() Results Between CPU and GPU
Segmentation Fault in torch.lu_unpack() with bfloat16 Tensor
`torch.cumsum()` on bfloat16 tensor with dtype=torch.int8 produce inconsistent results between CPU and GPU
torch.linalg.lstsq generating different solutions on CPU and GPU
torch.digamma produces inconsistent result between cpu and gpu for 0.00615 (float16)
torch.nansum does not work with complex numbers on CPU
`torch.nn.LocalResponseNorm` raises `RuntimeError` on CPU for float16 inputs (parity issue with CUDA)
torch.nn.Linear shape mismatch with large/incompatible inpu
torch.index_select out-of-bounds triggers inconsistent runtime errors
`torch.addr`: CPU/GPU Inconsistency with `int64` input and `float16` vector (CPU returns `inf`, GPU returns valid values)
torch.angle cuda version not implemented for bfloat16 and half (float16)
torch.floor_divide not triggering ZeroDivisionError on GPU
torch.nn.AvgPool2d works with long on cpu but not gpu
Silent Heap Buffer Overflow in standard ops (add, pinv) via implicit view resizing
`torch.clip` has checks for float16 scalar overflow on CPU but not on GPU
`torch.nn.PixelUnshuffle` throws RuntimeError: Stride calculation overflowed on CUDA with empty input and large downscale factor (CPU succeeds)
`torch.linalg.cholesky_ex`: Inconsistency between CPU (Inf) and CUDA (NaN) for infinite input, with silent NaNs on CUDA
`torch.nn.functional.PoissonNLLLoss` returns `NaN` imaginary part on CPU (vs 0 on GPU) for overflowing complex inputs
`torch.nn.functional.pdist` (p=0) with `inf` inputs produce wildly different values
torch.ldexp raises NotImplementedError for a bool exponent (eager), while torch.compile and the decomposition accept it
torch.distributions.Cauchy.log_prob returns -inf (and NaN gradients) for large |value - loc| / scale because the square overflows
torch.distributions.Normal.log_prob returns NaN for very large or very small scale because scale**2 overflows/underflows
torch.distributions.Laplace cdf/icdf return a zero gradient at the median (value == loc, p == 0.5)
torch.distributions.Laplace.icdf returns -inf (and loses accuracy) for small tail probabilities
torch.special.bessel_j0 / bessel_j1 / bessel_y0 lose ~9 significant digits in float64: the Cephes p1evl denominators drop their implied leading 1
[CPU] logit_backward and six activation backward kernels return a different gradient at NaN depending on tensor length
torch.minimum and torch.maximum return a zero whose sign depends on the tensor length
`torch.trace` raises `NotImplementedError` for `float16` input on CPU — no kernel registered for `Half`, no documented exclusion, NumPy handles it fine
torch.unique returns duplicate sub-tensors when the input contains NaN
torch.arctanh under torch.compile loses all precision for small inputs (returns 0.0)
torch.clamp with a bound unrepresentable in the tensor dtype: eager raises, torch.compile ignores the bound
TensorFlow
Deep Learning
61 bugs
38 Confirmed (Awaiting Fix)
12 Fixed
3 Pending
8 Rejected
38 Confirmed (Awaiting Fix)
12 Fixed
3 Pending
8 Rejected
`tf.linalg.svd` crashes on CUDA raising SIGABRT after facing OOM
`tf.experimental.numpy.cumsum` handles overflow inconsistently on CPU and GPU
`tf.linalg.matrix_rank` produces inconsistent output on CPU vs GPU with `tol=6`
`tf.nn.local_response_normalization` returns incorrect output
`tf.pow` returns inconsistent value on CPU vs GPU
`tf.nn.depthwise_conv2d` crashes with large `strides` values when ONEDNN is enabled
Inconsistent behavior for `tf.raw_ops.NotEqual` between CPU and GPU with non-broadcastable shapes
`tf.linalg.eigh` faces OOM on CUDA with 6D tensor and raises SIGABRT
`tf.linalg.lu_solve` returns inconsistent output from CPU vs GPU
`tf.raw_ops.MatrixDiag` crashes with `Aborted (core dumped)` with a 6D tensor as input on GPU
`tf.linalg.pinv` crashes with after raising SIGABRT with a 6D tensor due to OOM on CUDA
`tf.raw_ops.QuantizedConv2D` crashes with assertion failure instead of proper error handling
`tf.raw_ops.Mfcc` causes segmentation fault
Segmentation Fault in `tf.raw_ops.Conv3DBackpropInputV2`
Segmentation Fault in `tf.raw_ops.MaxPool` with oneDNN and `NCHW_VECT_C`
Segmentation Fault in `tf.raw_ops.QuantizeAndDequantizeV3`
Crash in `ResourceSparseApplyProximalAdagrad`
Crash in `SparseBincount` due to overflow
Floating Point Exception in `AvgPool3DGrad`
FPE in oneDNN `MaxPool3D` with Invalid Dimensions
`Segmentation Fault` in CropAndResizeGradBoxes
Segmentation Fault in `tf.raw_ops.HistogramFixedWidth`
Crash in `tf.raw_ops.ResourceSparseApplyAdagradDA`
Segmentation Fault in `tf.raw_ops.ResourceSparseApplyMomentum`
Fatal crash in `tf.raw_ops.TensorScatterMin` with malformed indices
Fatal crash in `tf.raw_ops.TensorScatterMax` with malformed indices
Inconsistent Error Handling in `tf.raw_ops.SparseSegmentSumGradV2` Between CPU and GPU Implementations
Inconsistent Error Handling in `tf.raw_ops.SparseToDense` Between CPU and GPU Implementations
Crash in `tf.raw_ops.BiasAdd` when executing on GPU
Crash in `tf.raw_ops.MaxPoolGradGradWithArgmax` when executing on GPU
`tf.conv2d_backprop_filter_v2` Fails with "no registered kernels" Error
Significant Discrepancy in `tf.linalg.triangular_solve` Results Between CPU and GPU
`tf.linalg.lstsq` producing outputs with large inconsistencies between CPU and GPU with `float32` tensors
`tf.math.floormod` not throwing `Integer division by zero` error on GPU for tensor of int64 dtype
Inconsistency (CPU vs. GPU) when casting NaN to int32 across multiple APIs
`tf.math.unsorted_segment_max` fails to propagate NaNs on GPU (inconsistent with CPU)
`tf.eye` with `TF_ENABLE_ONEDNN_OPTS=1` causes crash (Aborted/SegFault) on large `num_columns` instead of OOM
`tf.raw_ops.ParallelConcat` crashes (`SIGABRT`) due to integer overflow in shape calculation
`tf.raw_ops.AudioSpectrogram` Segfaults with negative `window_size` and `stride`
`tf.raw_ops.ImmutableConst` causes Segfault (`SIGSEGV`) with large shape dimensions
`tf.experimental.numpy.amax` crashes (`SIGABRT`) with specific int64 overflow in shape
`tf.io.encode_png` crashes process (`SIGABRT`) with `CHECK` failure when input has 0-dimension
Aborted (core dumped) in `tf.linalg.det/slogdet/logdet/cholesky/inv`
segmentation fault when tf.histogram_fixed_width receives large `value_range` and `nbins` on CPU mode
`tf.math.argmax` throws `InvalidArgumentError` with valid `axis` of `int16` dtype
Inconsistent Error Handling in `tf.raw_ops.SparseSegmentSqrtNGradV2` Between CPU and GPU Implementations
`tf.nn.conv2d_transpose` crashes with "Illegal instruction (core dumped)"
Inconsistent Results Between CPU and GPU for `tf.math.pow`
Discrepancy in `tf.experimental.numpy.nansum` with `dtype=float16` on CPU vs GPU
Misleading output when an invalid encoding Parameter Passed to tf.compat.as_str_any
tf.compat.dimension_value returns a Tensor when given a Tensor, despite documentation stating it returns only int or None.
tf.compat.forward_compatible does not validate year/month/day ranges as documented
When a single tensor and out of range axis is passed to tf.concat it silently succeeds
CPU/GPU inconsistency when casting bfloat16 to int8 for the api tf.experimental.numpy.asarray
Unclear Error Message in tf.raw_ops.QuantizedMaxPool with oneDNN Enabled
Discrepancy in Error Handling for `tf.gather` on CPU vs GPU Leading to Undefined Behavior
tf.raw_ops.SparseFillEmptyRowsGrad CPU-GPU inconsistency for out-of-range reverse_index_map
`tf.experimental.numpy.ones` crashes (segfault/abort) when INT64_MAX is passed as shape
`tf.experimental.numpy.var` Causes `SIGABRT` on Empty Tensors with Large Dimensions
`tf.math.reduce_variance` causes `SIGABRT` on large shapes containing zero
`tf.image.non_max_suppression` causes `SIGABRT` (Check Failure) with negative `max_output_size`
JAX
Autodiff / ML
38 bugs
17 Confirmed (Awaiting Fix)
17 Fixed
2 Pending
2 Rejected
17 Confirmed (Awaiting Fix)
17 Fixed
2 Pending
2 Rejected
betaln(a, b) is too large by exactly min(a, b) when min(a, b)/max(a, b) is subnormal
jnp.histogram assigns values lying exactly on an interior bin edge to the wrong bin
jax.scipy.stats.pareto returns -inf/inf/nan at the support boundary when scale is small relative to loc (support mask computed on unstandardized x)
jax.scipy.stats.expon.logcdf returns -inf for small positive x (log1p(-sf) cancellation)
jax.scipy.stats.laplace.cdf gives NaN gradients in the tails (unselected branch overflows)
jax.jit: negative static_argnums is not applied when the argument is passed by keyword (static_argnames not inferred)
jax.scipy.special.zeta: the custom_jvp rule returns a different primal value than the function (nan vs finite for s < 1)
jax.scipy.stats.poisson.entropy returns NaN gradients for large mu (and small mu in float32)
jax.scipy.special.beta returns nan / wrong sign at multiple defined points (negative integers, ±inf, -0)
jax.scipy.special.i1 overflows to inf for large finite float64 input
jax.scipy.stats.uniform.cdf/pdf/logpdf silently map NaN input to finite values (pdf(nan) == 1.0)
jax.scipy.stats.vonmises.pdf returns nan at kappa=0.0
jax.scipy.special.rel_entr(0.0, nan) returns inf instead of nan — SciPy compatibility mismatch
jax.scipy.special.kl_div returns nan instead of inf at x=inf, y=0.0
jax.scipy.stats.cauchy.ppf/isf: no quantile-domain handling — ppf(0) returns large finite value (wrong sign in float32), q outside [0,1] returns garbage
jax.scipy.stats.chi2.logpdf returns NaN at x=0 and x=inf when df=2 (inconsistent with gamma.logpdf)
jax.scipy.stats.dirichlet.logpdf returns -inf at simplex boundary where density diverges to +inf (alpha<1)
jax.scipy.stats.multinomial.logpmf/pmf return -inf for all rows of a batched input (sum reduces over all axes, not axis=-1)
jax.scipy.special.erfcx returns spurious 0.0 in a band near x≈26.6 (should be ~0.0212; scipy/torch agree)
`jax.scipy.special.xlogy` and `xlog1py` return `0.0` instead of `nan` when `x=0` and `y=nan`
`jnp.abs(inf + nan*j)` returns `nan` — violates IEEE 754 / C99 "infinity dominates" rule for complex magnitude
`jax.scipy.special.digamma(0.0)` returns `nan` — should be `-inf`
jax.scipy.special.loggamma returns nan instead of inf at x=0.0 and x=-0.0
jax.lax.zeta gives wrong results for negative non-integer q (Euler-Maclaurin breakdown)
jax.Array.sum silently ignores the `initial` argument
jax.fwd_and_bwd silently discards keyword arguments
`lax.pmean` with unequal `axis_index_groups` divides every group by the size of the first group
jax.scipy.special.gammaln(-inf) returns +inf instead of -inf — sign-wrong at infinity edge
jax.scipy.special.rel_entr / kl_div lose all precision for p close to q and return nan for large p (p*log(p) - p*log(q))
jax.scipy.special.gammaincc returns values far outside [0, 1] for large a (continued fraction stops unconverged)
jax.random.poisson: samples for large lam (>~2e5) have the wrong variance (+46% at lam=1e8)
jax.scipy.stats.t.logpdf is wrong for large df (lgamma(df/2) - lgamma((df+1)/2) cancels): logpdf(0, df=1e7) = 0 in float32
jnp.linalg.eigh silently accepts (…, 1, n) / (…, n, 1) input and returns the decomposition of a broadcast n x n matrix
jax.scipy.special.expn(n, 0) returns 1/n instead of 1/(n-1) for n >= 2
jnp.floor_divide / divmod return a quotient larger than x1/x2 for some float16 and bfloat16 inputs
jax.scipy.special.expi(0.0)` returns nan instead of -inf
jax.scipy.stats.poisson.logpmf` applies the integrality test to `k` instead of `k - loc`
`jnp.add.reduceat` and `jnp.multiply.reduceat` ignore the requested accumulator dtype
vLLM
LLM Serving
34 bugs
20 Confirmed (Awaiting Fix)
9 Fixed
5 Pending
20 Confirmed (Awaiting Fix)
9 Fixed
5 Pending
[Bug]: gigachat3, granite and granite-20b-fc tool parsers lose tool calls when the whole message arrives in a single streaming delta
[Bug]: Reasoning parsers mishandle markers in multi-token streaming deltas
fix(logprobs): guard convert_ids_list_to_tokens against out-of-range ids
[Bugfix] Parse human-readable ints with Decimal to avoid off-by-one
[Bugfix] Avoid an IndexError for trailing dotted CLI args
[Bugfix][Reasoning] Olmo3: keep literal think tags in streamed content
[Bugfix] Let an optional Literal flag accept the None it advertises
[Bugfix][Tool Parser] Return early for unlisted tool names in InternLM2 parser
[Bugfix] Invalidate LazyDict cache on __setitem__
[Bugfix][Core] Preserve None values in swap_dict_values
[Bugfix] Fix extreme case in pythonic parser argument json safety
[Bugfix][Structured Outputs] Reject empty `structural_tag` at request validation
[Bug]: xlam and hunyuan_a13b tool parsers drop the arguments of any tool call that follows an empty-argument call (streaming)
[Bugfix] Count reasoning tokens when the start token is absent
[Bugfix] reject empty prompt lists in CompletionRequest
[Bugfix][Frontend] Avoid mutating OpenAI request extra args
[Bugfix][Frontend] Count reasoning tokens for Harmony, DeepSeek-V3 and Step3 parsers
fix(reasoning): recompute ernie45 response-end index after stripping prefix
[Bugfix] Escape control characters in xgrammar choice grammar
[Bugfix] Make Lark-to-EBNF quote handling escape-aware
[Bugfix] Avoid mutating tool parameters in _get_tool_schema_defs
[Bugfix] Limit split to first marker in internlm2 tool parser
[Bug]: Qwen3MoeSparseMoeBlock rejects its documented 1D input
[Bug]: FourierRotaryEmbedding pads sin with 1 and corrupts unrotated channels
[Bugfix][LoRA] Skip no-LoRA tokens (index -1) in CPU torch bgmv ops
[Bug]: AudioSpec(target_channels=1) does not produce 1D mono output for single-channel 2D audio
[Bug]: vllm.utils.system_utils.suppress_stdout crashes when sys.stdout has no file descriptor
[Bug]: vllm.utils.system_utils.suppress_stdout silently suppresses stderr when sys.stdout is bound to fd 2
[Bug]: FilesystemResolver loads adapters outside VLLM_LORA_RESOLVER_CACHE_DIR (path traversal via LoRA name)
[Bug]: DeepSeek-V4 VL dummy image is not the worst case for multimodal profiling
[Bug]: ThinkingBudgetStateHolder maps every request to logits row 0 when spec decoding is on but a step has no drafts
[Bug]: ThinkingBudgetStateHolder.sync_batch leaks stale state on unidirectional moves
[Bug]: vLLM hangs indefinitely with low `num_gpu_blocks_override`
[Bug]: `block_size=8` triggers Triton CompilationError in FlexAttention kernel; other backends correctly reject
SGLang
LLM Serving
33 bugs
16 Confirmed (Awaiting Fix)
1 Fixed
15 Pending
1 Rejected
16 Confirmed (Awaiting Fix)
1 Fixed
15 Pending
1 Rejected
fix(completion): treat None suffix as empty in FIM prompt builder
[Fix] Keep normal text preceding the tool-call opener in streaming parsers
[Bug] 9 tool-call parsers drop arguments when one streaming increment has more than one complete call
[Bug] 8 tool-call parsers give wrong or missing calls in token-by-token streaming
fix(function_call): skip non-object entries in parse_base_json
[Bug] MiMo tool-call streaming: text after the last tool call is silently dropped; split bot-token flushes markup into content
fix(function_call): strip think tags split across stream chunks in Trinity
[Fix] Cohere Command4: strip echoed `<|START_THINKING|>` in streaming and flush the final text block
fix(function_call): handle non-finite numbers in tool-call argument coercion
[Bug] JIT DeepGEMM is enabled on SM121 (GB10 / DGX Spark): the SM120 exclusion checks `== 120` instead of the consumer-Blackwell family
[parser] Hold back harmony markers split across streaming chunks
fix: truncate xgrammar accepted_tokens in place on rollback
Apertus2509Detector (streaming) drops the second tool block and trailing text when they arrive in the same chunk`
[Ollama] Map options.seed to the sampling_seed field
Tool-call parsers lose or corrupt data at streaming chunk boundaries (multiple detectors)
[Fix] hunyuan: Parse tool arguments with JSON Schema type arrays
[Fix] function_call: Flush pending tool calls at end of stream in the Llama 3.2 detector
[Bugfix] Preserve single quotes inside JSON values in Llama32 streaming tool calls
Tool-call parsers lose or corrupt data at streaming chunk boundaries (multiple detectors)
[Bug] deepseekv32/deepseekv4 tool-call parsers drop calls when the streamed output has two chunks
[Bug] DeepSeekV31Detector.structure_info() omits `<|tool▁calls▁begin|>`, so its own detect_and_parse returns no tool call
[Bug] Glm4MoeDetector streaming emits invalid JSON when a non-string parameter's value is not valid JSON
[Bug] Glm4MoeDetector changes string-typed argument values that look like JSON ("true" → "True", "1.50" → "1.5")
[Bug] ChatCompletionRequest.set_json_schema mutates the caller's response_format schema (reused schemas change behaviour)
[Bug] cohere_command4 detector (streaming) drops the tool call when text and the whole action block arrive in one chunk
[Bug] deepseekv31 detector (streaming) drops the text that shares a delta with the start of a tool call
[Bug] deepseekv31 detector (streaming): greedy name regex drops a call and attaches the next call's arguments to it
[Bug] hunyuan detector (streaming): control characters in string arguments produce invalid, corrupted JSON
[Bug] trinity detector removes `<think>` / `</think>` from inside tool-call arguments
[Bug] Ollama streaming endpoints mangle the output when `--incremental-streaming-output` is on
[Bug] `get_hf_text_config` raises AttributeError when `thinker_config` holds a dict `text_config`
[Bug] HarmonyParser streaming: arguments of a tool call on the analysis channel are emitted as reasoning
[Bug] XGrammarGrammarBackend._sanitize_structural_format` skips `optional`, `star`, `plus`, `repeat`, `dispatch`, `token_dispatch` and `token_triggered_tags`, so a `null` `json_schema` inside them is rejected
ArangoDB
Graph Database
24 bugs
1 Confirmed (Awaiting Fix)
10 Fixed
12 Pending
1 Rejected
1 Confirmed (Awaiting Fix)
10 Fixed
12 Pending
1 Rejected
AQL `SUBSTITUTE()` returns `null` for an empty input string
`SUBSTITUTE` returns null and an ICU warning for an empty input string
`TO_NUMBER` converts a valid subnormal decimal string to zero
`APPEND(null, [])` returns null instead of the other input array
REVERSE corrupts a single astral Unicode character
`FIND_FIRST` and `FIND_LAST` fail to find present control characters
`REGEX_MATCHES` reports a match for an empty string that cannot match
AT LEAST (<n>) (<expression>) is true even when n is larger than the array's length
Persistent index with deterministic SORT and offset LIMIT incorrectly returns an empty result
ArangoDB DATE_ISO8601() applies the sign of a negative UTC offset only to its hours, not its minutes
ArangoDB `NGRAM_MATCH()` counts one attribute n-gram many times (stale DP diagonal)
ArangoDB AQL `SORTED()` replaces distinct elements that compare equal
`JSON_PARSE` exponent numbers can read beyond the input string
`JSON_PARSE` can misround fractions and large integer tokens
AQL `use-indexes` uses a sparse index for a range condition that can match `null` when an `OR` branch proves non-null, dropping rows
AQL `SUBSTITUTE()` with several search values skips a match that starts right where the previous replacement ended
AQL arithmetic on 64-bit integers from documents loses precision; the same expression on literals does not
AQL `remove-redundant-calculations` treats `(a ? b : c) ? d : e` and `a ? b : (c ? d : e)` as the same calculation
`POST /_api/import` stores duplicate attribute names from document and tabular input
`/_api/import` with `type=documents` and `complete=true` commits a partial import when a line is not valid JSON
AQL `COLLECT` served from a persistent index merges the array values `[]` and `[null]` into one group
AQL ternary operator evaluates an untaken operand when it contains an array expansion with a subquery
`DATE_ROUND` with a large positive amount terminates the server with SIGFPE
`DATE_ROUND` assigns pre-1970 timestamps to the following bucket
cronie
System Daemon
13 bugs
12 Confirmed (Awaiting Fix)
1 Fixed
12 Confirmed (Awaiting Fix)
1 Fixed
Support input from a pipe for `crontab -T -`
`anacron -T` error stream contradicts its manual
anacron: START_HOURS_RANGE, RANDOM_DELAY and PREFERRED_HOUR are matched by prefix, so other variables such as START_HOURS_RANGE_OLD are rejected or silently change scheduling
anacron: misspelled period macros such as `@monthlyXYZ` are accepted as `@monthly`, and a job with an unknown macro runs although anacron says it is skipped
anacron: START_HOURS_RANGE hours are not range-checked (and overflow `int`), so invalid ranges pass `anacron -T` and silently stop all jobs
cronnext: `-t` and `-q` parse the epoch time with `atoi()`, so times from 2038-01-19 on are truncated (Y2038)
@daily`/`@midnight` and `@hourly` do not set `DOM_STAR`/`DOW_STAR`, unlike their documented equivalents (cf. #52)
`cronnext` ignores a stepped day-of-week / day-of-month field
reversed ranges are accepted but never match
`cronnext` ignores `CRON_TZ`
cronnext reports a time one hour late when skipping to a month with a different DST state
cronnext can skip an entire matching day across a DST spring-forward transition
anacron: a large `RANDOM_DELAY` value crashes anacron with SIGFPE or makes it loop forever
PostgreSQL
Database Engine
10 bugs
2 Fixed
8 Pending
2 Fixed
8 Pending
pg_backup_start accepts a newline in the label, and the backup will not start
first_value/last_value/nth_value return NULL under EXCLUDE TIES when the current row is outside its frame
first_value/last_value/nth_value return NULL under EXCLUDE TIES when the current row is outside its frame
an empty JSON_OBJECT rejects the ON NULL and WITH UNIQUE KEYS clauses that its documented syntax allows
transaction_timestamp() and statement_timestamp() differ for a parameterized first statement, which the documentation says return the same value
has_language_privilege returns true for a nonexistent language OID when the current user is a superuser, where a named non-superuser gets null
avg(double precision) overflows on cancelling finite inputs whose mean is representable
jsonb_object_agg_unique_strict silently discards a JSONB null value
JSON_OBJECT with ABSENT ON NULL accepts a NULL key and an array, composite or JSON key for the omitted pair
uuidv7(interval) rejects timestamps 1 to 999 microseconds into the last millisecond its field can hold
Clang / LLVM
Compiler / Toolchain
9 bugs
6 Confirmed (Awaiting Fix)
1 Fixed
2 Pending
6 Confirmed (Awaiting Fix)
1 Fixed
2 Pending
[Clang] Assertion `is<TypeInfoLValue>() && "not a type_info lvalue"' failed with alignment builtins
clang `__is_bitwise_cloneable` hangs the compiler on nested structs (exponential, un-memoized recursion)
clang `model` attribute silently ignored when the variable was referenced before its definition
Clang misassigns homogeneous vector aggregate arguments for SysV `vectorcall`
clang `__attribute__((vectorcall))` crashes on any `long double` parameter (x86_64 Linux)
[clang] `constexpr` `__builtin_align_down/up` on a pointer into an array reads the original element
`__builtin_convertvector` from `__bf16` to `long double` vectors gives wrong values
`__builtin_convertvector` to a `bool` vector folds 0.5 to `false` in a constant expression
`nonnull` whose every index is rejected becomes "all pointers" and changes code
GCC
Compiler / Toolchain
5 bugs
3 Confirmed (Awaiting Fix)
1 Pending
1 Rejected
3 Confirmed (Awaiting Fix)
1 Pending
1 Rejected
__builtin_is_virtual_base_of reports a non-virtual base reached through a virtual base as virtual
the manual's equivalent expression for __builtin_stdc_rotate_left/right is wrong for non-power-of-two _BitInt widths
implicitly zero-initialized hardbool objects hold zero bits instead of the false representation
the GCC manual says , ## __VA_ARGS__ drops the comma for an empty argument; the CPP manual and GCC keep it
__builtin_bitreverseg on a dependent C++ operand gives an ICE in mangle_decl
OpenSSH
Security / Network
5 bugs
1 Fixed
4 Pending
1 Fixed
4 Pending
ssh -G prints EscapeChar ^A as \^A which does not parse back
ssh -G shows some LocalForward entries as DynamicForward
ssh-keygen -s signs certificates with duplicate options, which sshd rejects
StreamLocalBindMask ignores Host and Match blocks, and the last value wins
ssh-keygen -F and -R miss a hashed known_hosts entry when the host name's case differs