-
Notifications
You must be signed in to change notification settings - Fork 714
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
- Status: Open.#1261 In fla-org/flash-linear-attention;
- Status: Open.#1255 In fla-org/flash-linear-attention;
[Bug] NPU triton-ascend kernels silently mismatch on unaligned K/V/T (CANN leftover DMA aliases the next row)
ascend-npuAscend NPU (triton_ascend) relatedAscend NPU (triton_ascend) relatedbugSomething isn't workingSomething isn't workingStatus: Open.#1251 In fla-org/flash-linear-attention;[Bug] KDA safe_gate fast path can make prefix outputs and gradients depend on future gates
needs-verificationLacks real execution evidence (CI skipped / no before-after data)Lacks real execution evidence (CI skipped / no before-after data)Status: Open.#1240 In fla-org/flash-linear-attention;- Status: Open.#1234 In fla-org/flash-linear-attention;
[RFC] Add Graph Capture Support for FLA Prefill and Training Paths on Ascend NPUs
ascend-npuAscend NPU (triton_ascend) relatedAscend NPU (triton_ascend) relatedenhancementNew feature or requestNew feature or requestStatus: Open.#1233 In fla-org/flash-linear-attention;[Bug] tl.dot(..., allow_tf32=False) returns NaN/Inf in recompute_w_u_fwd_kernel_npu with fully-finite bf16 inputs
ascend-npuAscend NPU (triton_ascend) relatedAscend NPU (triton_ascend) relatedbugSomething isn't workingSomething isn't workingStatus: Open.#1215 In fla-org/flash-linear-attention;[RFC] Graph capture support for prefill and training: GDN/KDA chunk and convolution paths
enhancementNew feature or requestNew feature or requestStatus: Open.#1155 In fla-org/flash-linear-attention;[Bug] High Local Disk IOPS and Frequent File Open Operations with FLA 0.5.1 During Multimodal Model Training
bugSomething isn't workingSomething isn't workingStatus: Open.#1072 In fla-org/flash-linear-attention;[RFC] Roadmap: NPU support for flash-linear-attention
enhancementNew feature or requestNew feature or requestStatus: Open.#942 In fla-org/flash-linear-attention;[Feature Request] Introducing an input buffer for
final_stateinfused_recurrent_gated_delta_rule_fwdenhancementNew feature or requestNew feature or requestStatus: Open.#872 In fla-org/flash-linear-attention;[Feature Request] Implement tensor inverse
enhancementNew feature or requestNew feature or requestStatus: Open.#680 In fla-org/flash-linear-attention;