Repository Issue Activity (beta)

#430 [nv_dev][SM120] k_grouped_fp8_gemm_nt_contiguous returns non-deterministic wrong results — sm120 kernel is missing the tensormap drain fixed for SM90 in #343

#425 sm121a: `paged_mqa_logits` varlen path produces illegal memory access; uniform path does not scale past batch 1 at long context

#417 [Regression] 8b1392b removes SM12x pure-fp8 1d1d kernels and aliases fp8_gemm_nt to the fp4 dispatcher (silent corruption on GB10)

#405 CUDA Illegal Instruction in test_k_grouped_gemm_contiguous

#414 [Bug] `pack_ue8m0_to_int` is not CUDA Graph capture safe

#412 get_env<int> returns uninitialized data for malformed values

#410 [RFC] Chunk-ready symmetric-buffer protocol for streaming MegaMoE inputs

#385 Origin & calibration methodology of the three cost-model constants in sm120.hpp

#402 Measured data point on fixed-tile vs shape-adaptive GEMM: where a bit-exact fixed-tile kernel stops (and why we couldn't fix it)

#392 [nv_dev] SM120: MQA logits kernels fail test_attention self-consistency (fp8: nondeterministic/race) and accuracy (mxfp4 H=64 D=128)

#391 Question About deepGEMM contributor and Collaboration

#322 `get_paged_mqa_logits_metadata()` does not work with large batch size due to excessive static shared memory impl

#365 [BUG] Non-deterministic numerical problem of k_grouped_fp8_gemm_nt_contiguous

#374 [DSv4] IS W4A16(mxfp4, act bf16) on the planning list?

#372 transform_sf_into_required_layout: no SM120/121 (arch_major=12) branch for (gran_mn=1, gran_k=32) NVFP4 expert scales -- blocks DeepSeek-V4 weight loading

#366 [Feature] Will DeepGEMM support more advanced epilogue fusion for GEMM kernels?

#370 Incompatibility between DeepEP V2 ElasticBuffer non-expanded output and DeepGEMM masked/contiguous grouped GEMM input format

#336 Question about Nsight Compute results for sm100_fp8_fp4_mega_moe_impl on B300/Blackwell

#368 question about tmem barrier

#367 question about fp8fp4 mega moe weights load

#339 [BUG] Mega MoE kernel中的小错误

#363 00

#359 [BUG] NVRTC build fails for `sm90_tf32_hc_prenorm_gemm` — `cudaGridDependencySynchronize` undefined (PDL intrinsic not available under NVRTC)

#358 [Feature request] Documented clangd setup for CUDA kernel indexing

#356 Does Mega MoE support RTX 50-series GPUs, and can it work in a PCIe-only setup without NVLink?