|
|
@@ -4,6 +4,15 @@
|
|
|
# same image, same NVFP4 checkpoint, only the speculative model differs.
|
|
|
# (This recipe = glm53-flash-aiden-4x with the drafter swapped.)
|
|
|
#
|
|
|
+# PARKED 2026-09-04: with the MXFP8→bf16 dequantized lab drafter the target
|
|
|
+# loads clean and the drafter loads (no missing params), but the FIRST draft
|
|
|
+# pass at warmup dies in _project_context_kv: "mat1 and mat2 same dtype —
|
|
|
+# BFloat16 vs Float8_e4m3fn" (the fused KV-proj weight ends up fp8 at
|
|
|
+# runtime — the NVFP4 target's quant config bleeds into the draft's linears,
|
|
|
+# and the lab drafter's layout doesn't survive that path; incoai's does).
|
|
|
+# The lab drafter belongs to the lab's own vllm fork (newer dflash + aux-
|
|
|
+# capture). Revisit with a lab-fork image build.
|
|
|
+#
|
|
|
# aiden's production 4-node profile, verbatim translation of
|
|
|
# https://aidenle.com/recipes/glm-5-3-flash-4x-dgx-spark into our 4x launcher
|
|
|
# shape. Reference numbers (his, single stream, idle cluster, thinking off):
|
