forked from danveloper/flash-moe
-
Notifications
You must be signed in to change notification settings - Fork 1
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
cuda-qwen36: port HTTP/SSE serve + tool calling from infer.cu (Qwen3.5-397B path)
port:cuda-qwen36Affects the CUDA Qwen3.6 port (cuda-qwen36)Affects the CUDA Qwen3.6 port (cuda-qwen36)propagate:cuda-qwen36This issue still needs the fix on cuda-qwen36This issue still needs the fix on cuda-qwen36reliabilityBlocks the reliability bar (agent loop + small eval)Blocks the reliability bar (agent loop + small eval)tool-callingTool calling, JSON output, agent loopTool calling, JSON output, agent loopStatus: Open.#5 In ssubbotin/flash-moe;Tokenizer subprocess: kimi_tokenize.py is generic, but each port carries its own copy
infraBuild, CI, repack scripts, tokenizer subprocessBuild, CI, repack scripts, tokenizer subprocessport-onlyIntentionally branch-local; do NOT propagateIntentionally branch-local; do NOT propagateport:cuda-kimiAffects the CUDA Kimi K2.6 port (cuda-kimi)Affects the CUDA Kimi K2.6 port (cuda-kimi)port:cuda-qwen36Affects the CUDA Qwen3.6 port (cuda-qwen36)Affects the CUDA Qwen3.6 port (cuda-qwen36)Status: Open.#4 In ssubbotin/flash-moe;Verify embed >2GB pread fix is not latent on cuda / cuda-qwen36 / rocm
correctnessOutput drifts from oracle / referenceOutput drifts from oracle / referenceport:cudaAffects the generic CUDA backend (cuda)Affects the generic CUDA backend (cuda)port:cuda-qwen36Affects the CUDA Qwen3.6 port (cuda-qwen36)Affects the CUDA Qwen3.6 port (cuda-qwen36)port:rocmAffects the ROCm backend (rocm)Affects the ROCm backend (rocm)propagate:cudaThis issue still needs the fix on the CUDA backendThis issue still needs the fix on the CUDA backendpropagate:cuda-qwen36This issue still needs the fix on cuda-qwen36This issue still needs the fix on cuda-qwen36propagate:rocmThis issue still needs the fix on rocmThis issue still needs the fix on rocmStatus: Open.#3 In ssubbotin/flash-moe;Metal 2-bit experts break JSON output (quotes → backslashes)
port:metalAffects the Metal backend (main)Affects the Metal backend (main)propagate:metalThis issue still needs the fix on the Metal backendThis issue still needs the fix on the Metal backendreliabilityBlocks the reliability bar (agent loop + small eval)Blocks the reliability bar (agent loop + small eval)tool-callingTool calling, JSON output, agent loopTool calling, JSON output, agent loopStatus: Open.#2 In ssubbotin/flash-moe;Kimi K2.6: hidden state collapses to zero between L1 and L5
correctnessOutput drifts from oracle / referenceOutput drifts from oracle / referenceport:cuda-kimiAffects the CUDA Kimi K2.6 port (cuda-kimi)Affects the CUDA Kimi K2.6 port (cuda-kimi)propagate:cuda-kimiThis issue still needs the fix on cuda-kimiThis issue still needs the fix on cuda-kimireliabilityBlocks the reliability bar (agent loop + small eval)Blocks the reliability bar (agent loop + small eval)Status: Open.#1 In ssubbotin/flash-moe;