|
|
@@ -1,3 +1,11 @@
|
|
|
+# NOTE: the lab's DFlash2 checkpoints are MXFP8-quantized (weight_scale
|
|
|
+# tensors) and their config names the arch DFlash2DraftModel — the lab image
|
|
|
+# registers the drafter under DFlashDraftModel and its loader has no MXFP8
|
|
|
+# scale support (ValueError: no parameter 'weight_scale'). So this lane uses
|
|
|
+# a dequantized bf16 copy (exact value reconstruction of the MXFP8 tensors)
|
|
|
+# packaged as local-inference-lab/GLM-5.3-Flash-DFlash2-bf16-lab (arch
|
|
|
+# DFlashDraftModel) in the shared NAS hub. The aiden-image lane uses the
|
|
|
+# sibling -bf16 repo (arch DFlash2DraftModel).
|
|
|
# GLM-5.3-Flash NVFP4 — local-inference-lab quant (vLLM) — 2× DGX Spark, TP=2.
|
|
|
#
|
|
|
# local-inference-lab mixed quant: NVFP4 routed experts (layers 3–44,
|
|
|
@@ -84,7 +92,7 @@ args = [
|
|
|
'{"enable_cutedsl_warmup":false,"enable_flashinfer_autotune":false}'""",
|
|
|
# MTP-3 draft from the MXFP8 layer-45 shard (needs modelopt.patch in image)
|
|
|
"--speculative-config", """
|
|
|
-'{"method":"dflash","model":"local-inference-lab/GLM-5.3-Flash-DFlash2","num_speculative_tokens":7}'""",
|
|
|
+'{"method":"dflash","model":"local-inference-lab/GLM-5.3-Flash-DFlash2-bf16-lab","num_speculative_tokens":7}'""",
|
|
|
"--language-model-only",
|
|
|
# MTP + CUDA graphs degrades OUTPUT QUALITY on the glm53 day-0 dev build
|
|
|
# (forum #381350/#44, CONFIRMED by A/B bench 2026-08-28: same seed/flags,
|
