3 people 3c051b52ed [Perf][SM70] Flash-Next: concurrent MTP4 verification batches (#1369) ...

Size the MTP draft MoE, shared gate and push all-reduce for up to 16 requests,
leave GDN and router batches above their row limits to cuBLAS instead of pieces,
add a fused dense MoE permute, and run HCX only for one request's rows. Also fix
the host-KV startup crash (slice output_gate to num_tokens) and copy every mapped
host KV run on the copy engine.

Co-authored-by: yangzhuxinyzx <Susannah_Carterect@gmx.ca>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

2026-11-05 22:45:28 +08:00

..

fixtures

…

gguf

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

glm5next

[Bugfix][KV] MLA specs carry the codec of their cache dtype (#1272)

2026-10-28 10:29:47 +08:00

language

…

multimodal

…

quantization

…

qwen3_5

…

qwen4_exp

[Perf][SM70] Flash-Next: concurrent MTP4 verification batches (#1369)

2026-11-05 22:45:28 +08:00

__init__.py

…

registry.py

…

test_adapters.py

…

test_bailing_moe_v3.py

…

test_deepseek_v4_fp8_capability.py

…

test_deepseek_v4_mega_moe.py

…

test_deepseek_v4_pipeline.py

…

test_deepseek_v4_rocm_swa_ragged.py

…

test_deepseek_v4_sm70_routes.py

…

test_deepseek_v4_sm70_sparse_policy.py

…

test_deepseek_v41_attention_ops.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_contract.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_dspark.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_engram.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_fp8_linear.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_graph_metadata.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_norm.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_quant_rope.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_query_graph.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_deepseek_v41_runner_state.py

[Model] DeepSeek-V4.1-Flash on 4x V100: 512K packed KV, faster prefill, API launch (#1277)

2026-10-27 23:33:35 +08:00

test_dflash_context_slots.py

…

test_gguf_download.py

…

test_initialization.py

…

test_language_model_cache_is_weak.py

…

test_mtp_skip_checkpoint_weight.py

…

test_oot_registration.py

…

test_pixtral_rope_compat.py

…

test_ple_optional_cuda_import.py

…

test_qwen4_ple_placeholder.py

…

test_registry.py

…

test_terratorch.py

…

test_transformers.py

…

test_utils.py

…

test_vision.py

…

utils.py

…