File size: 720 Bytes
80692f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | """Canonical shape sweep for KDA forward (chunk form).
Mix of:
- short-context training-step scale (T=1024)
- mid-context (T=2048) which is the headline benchmark
- long-context that stresses the inter-chunk recurrence (T=4096)
- thin-batch decode-style (B=1, T=2048, fewer heads)
Constraints:
- T % chunk_size == 0 (chunk_size = 64)
- K, V are the per-head channel dims; KDA in Kimi Linear uses K=V=128
"""
SHAPES = [
{"B": 2, "T": 1024, "H": 8, "K": 128, "V": 128, "CHUNK_SIZE": 64},
{"B": 2, "T": 2048, "H": 8, "K": 128, "V": 128, "CHUNK_SIZE": 64},
{"B": 1, "T": 4096, "H": 8, "K": 128, "V": 128, "CHUNK_SIZE": 64},
{"B": 1, "T": 2048, "H": 4, "K": 128, "V": 128, "CHUNK_SIZE": 64},
]
|