whn09 e7cb740b97 Neuron attention: decode fp4 KV records in float arithmetic (Vector engine), not integer bit ops (GpSimd) ... The compressed-KV scopes are fp4 records; their decode was nibble masks and shifts plus the integer e2m1 / e4m3 decoders on GpSimd (three such chains were the largest cost of the ratio-2 attention graph, device profile bench72). Now: nibbles from floor(b / 16), e2m1 from a piecewise-linear map of the magnitude code, scales through the hardware fp8 conversion. Bit-identical to the integer decode for every byte (host check). bench73 (vllm-sonnet 512, TP=64, nextn=5): TPOT 243.5 -> 228.6 ms, TTFT 4.84 -> 4.35 s, step 699 -> 666 ms, tokens per step identical. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>		2026-10-26 08:18:39 +08:00
..
nki	Neuron MoE: skip unrouted experts inside the fp4 kernels at decode sizes (DEEPINFER_NEURON_GATED_EXPERTS, default on)	2026-10-26 06:36:48 +08:00
vendor	Neuron engram shared tables: one file for the whole group (#28)	2026-10-24 07:08:37 +08:00
__init__.py	Add AWS Trainium (Neuron) platform, basic mode — part 1	2026-10-20 18:34:51 +08:00
fp8_guard.py	Neuron: tiny model on a NeuronCore matches the reference (issues #1, #2, #4)	2026-10-21 13:22:59 +08:00
fp8.py	Neuron: full DeepSeek-V4.1-Flash load + decode region graphs at EP=64 (issues #21, #25)	2026-10-23 01:39:22 +08:00
neuron_compile.py	Neuron: separate state regions, chunk-wide indexer context (issue #26)	2026-10-23 03:46:56 +08:00
NEURON_PORT.md	Neuron: tiny model on a NeuronCore matches the reference (issues #1, #2, #4)	2026-10-21 13:22:59 +08:00
nki_attention.py	Neuron attention: decode fp4 KV records in float arithmetic (Vector engine), not integer bit ops (GpSimd)	2026-10-26 08:18:39 +08:00
nki_gemm.py	Neuron MoE: skip unrouted experts inside the fp4 kernels at decode sizes (DEEPINFER_NEURON_GATED_EXPERTS, default on)	2026-10-26 06:36:48 +08:00
nki_hc.py	Neuron: per-layer region graphs + vllm-neuron#71 NKI kernels (issues #10, #11, #13, #25)	2026-10-22 17:36:54 +08:00
nki_norm.py	Neuron: full DeepSeek-V4.1-Flash load + decode region graphs at EP=64 (issues #21, #25)	2026-10-23 01:39:22 +08:00
nki_topk.py	Neuron: nkilib rotational_topk for the sampler; single-host shared engram tables (#27)	2026-10-23 07:23:32 +08:00
op_census.py	Neuron: compiled log-softmax + greedy argmax in the sampler; eager op census (#27)	2026-10-23 09:42:08 +08:00
regions.py	Neuron DSpark: 6-row steps keep only coarse region splits (8 graphs a layer instead of 12)	2026-10-25 00:31:45 +08:00
static_moe.py	Neuron MoE: skip unrouted experts inside the fp4 kernels at decode sizes (DEEPINFER_NEURON_GATED_EXPERTS, default on)	2026-10-26 06:36:48 +08:00
transfer_to_neuron.py	Neuron: static decode graphs, static-shape MoE, trn2 e4m3 fp8 (issues #9, #12, #24)	2026-10-22 11:59:37 +08:00