Initial K3 snapshot: 0.5B KDA/MLA/MoE train path

Standalone tree split from LLMRL/projects/kda. Includes Triton dt_bias
backward fix, train_k3 --preset 0.5b, SFT, Docker runtime, and tests.
This commit is contained in:
dela
2026-08-25 14:43:17 +08:00
commit 584f7e9e73
140 changed files with 21592 additions and 0 deletions
+13
View File
@@ -0,0 +1,13 @@
"""Context-parallel stubs. Pass ``cp_context=None`` (the default)."""
class FLACPContext:
cu_seqlens = None
cu_seqlens_cpu = None
def build_cp_context(*args, **kwargs):
raise RuntimeError("Context parallel is not included in the vendored KDA kernels")
__all__ = ["FLACPContext", "build_cp_context"]
+11
View File
@@ -0,0 +1,11 @@
"""Context-parallel hooks referenced by KDA fwd/bwd. Not implemented here."""
def _cp_unsupported(*args, **kwargs):
raise RuntimeError("Context parallel is not included in the vendored KDA kernels")
chunk_gated_delta_rule_fwd_h_pre_process = _cp_unsupported
compress_h0 = _cp_unsupported
chunk_gated_delta_rule_bwd_dhu_pre_process = _cp_unsupported
expand_h0 = _cp_unsupported