ATOM-native Kimi-K3 DSpark E4M3FN
This checkpoint is the ATOM-only, gfx950-native form of the Phase 1 FP8 PTPC
model. It stores global unshuffled tensors, pre-merges gate_up_proj and
fused_qkv_a_proj, and adds a BF16 context_kv_proj. Tensor-parallel slicing
and AITER preshuffle still happen at load time.
It requires model_type=atom_k3_dspark / AtomK3DSparkModel support from ATOM
and is intentionally not compatible with stock Transformers, vLLM, or SGLang.
TP8 performance
Measured on 8×MI350X (gfx950) with concurrency 8, 80 prompts, fixed 1024-token inputs, 512-token outputs, 7 speculative tokens, and FP8 KV cache.
| Metric | Portable Phase 1 FP8 PTPC | ATOM-native E4M3FN |
|---|---|---|
| Output throughput | 667.79 tok/s (3-run median) | 665.48 tok/s (1 run) |
| Mean TPOT | 9.95 ms | 10.59 ms |
| Speed acceptance rate | 54.85% | 52.59% |
| Accept length | 4.840 tokens/forward | 4.681 tokens/forward |
| Effective DSpark forwards/s | 138.0 | 142.2 |
The native checkpoint completed TP8 startup, runtime slicing, AITER preshuffle, graph capture, and real decode without unloaded or dropped parameters. Its steady-state throughput is within 0.35% of the portable Phase 1 result; the native format primarily provides a fail-closed ATOM ABI, offline projection merging, and the dedicated context-KV path.
ATOM support is available on the
lumen-rl-atom-native-dspark
branch. The exporter is on Lumen-RL
dev/atom-native-dspark.
- Downloads last month
- -