Upload verified Muse Spark DSpark step-2000 checkpoint

export_manifest.json

@@ -1,12 +1,12 @@
 {
   "format_version": 1,
-  "created_at_utc": "2026-09-09T23:27:42.194030+00:00",
-  "source_checkpoint": "/home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/evaluation/step900_e2e/native",
+  "created_at_utc": "2026-09-10T19:55:48.044685+00:00",
+  "source_checkpoint": "/home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/evaluation/checkpoint_20260910_195524/native",
   "target_model": "/home/inf-zixi/.cache/huggingface/hub/models--meta-models--Muse-Spark-1.2-820B-A42B-NVFP4/snapshots/b8ace9535a34b9c33a2ef5bfbc86189f56c73e2b",
   "source_config_sha256": "0cd8fa598e4939499bede4156d40c6abd1e945dc4b1c12bfb8d6442782525b9d",
-  "source_weights_sha256": "107ad8e99b66eab4a69bbebd6c23e1622ff71a6bfdd697c103e4939442e16bcf",
+  "source_weights_sha256": "6d0aa34c59ff1fc72dcbafb06271e892fa835747fe59dd07e85cba3d6217e759",
   "output_config_sha256": "588298a4cd6e5179467656644c33d29c31f33f4c951f5c26d2b43aaa390b9af7",
-  "output_weights_sha256": "9f84c053c3f323e17156c3b7eee3004508c28288b9673430613d2fdc3fa65242",
+  "output_weights_sha256": "0d6fd02243888d18e1495cc672c299d660849c390ffb6f6d8b872ac7208b4ada",
   "source_weight_bytes": 7292920442,
   "output_weight_bytes": 10604891514,
   "tensor_count": 74,
@@ -16,8 +16,8 @@
     "lm_head.weight"
   ],
   "training_state": {
-    "epoch": 0,
-    "local_step": 7200,
-    "global_step": 900
+    "epoch": 2,
+    "local_step": 1584,
+    "global_step": 2000
   }
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9f84c053c3f323e17156c3b7eee3004508c28288b9673430613d2fdc3fa65242
+oid sha256:0d6fd02243888d18e1495cc672c299d660849c390ffb6f6d8b872ac7208b4ada
 size 10604891514

train_command.txt

@@ -1,9 +1,9 @@
-# Timestamp: 2026-09-09T15:59:10.790877+00:00
-# Git SHA: d0ef02ae60143d0332171d75da1296a4009aa3df
+# Timestamp: 2026-09-10T15:32:19.131341+00:00
+# Git SHA: f71a86e0dbfd1e7801ecc225d9c3ca0423a52863
 # World size: 16
 # speculators: 0.8.0.dev226
 # vllm: 0.27.2rc1.dev107+gd6f17f3d5.precompiled
 # transformers: 5.14.1
 # torch: 2.13.0+cu130
 # compressed-tensors: 0.17.0
-scripts/train.py --speculator-type dspark --block-size 7 --markov-rank 512 --markov-head-type vanilla --enable-confidence-head --confidence-head-with-markov --from-pretrained /home/inf-zixi/speculator-training/muse_spark/training2/pilots_20260909/all_global/initial --verifier-name-or-path /home/inf-zixi/.cache/huggingface/hub/models--meta-models--Muse-Spark-1.2-820B-A42B-NVFP4/snapshots/b8ace9535a34b9c33a2ef5bfbc86189f56c73e2b --target-layer-ids 14 30 46 58 62 --reuse-last-hidden-state-as-aux --draft-vocab-size 202048 --mask-token-id 200018 --draft-attn-impl simple_flex_attention --loss-implementation fused --loss-fn '{"ce": 0.1, "tv": 1.8}' --dflash-decay-gamma 4.0 --per-position-loss-weight fixed-exp-decay --max-anchors 1024 --data-path /home/inf-zixi/speculator-training/muse_spark/training2/pilots_20260909/data --train-data-ratio 0.98 --total-seq-len 32768 --noise-std 0 --num-workers 5 --prefetch-factor 2 --hidden-states-dtype bfloat16 --hidden-states-backend mooncake --mooncake-master 10.78.208.151:50051 --mooncake-metadata-server P2PHANDSHAKE --mooncake-protocol tcp --mooncake-global-segment-size 0 --mooncake-local-buffer-size 8589934592 --vllm-endpoint http://r01-p01-dgx-c01:8000/v1,http://r01-p01-dgx-c02:8000/v1,http://r01-p01-dgx-c03:8000/v1,http://r01-p01-dgx-c05:8000/v1,http://r01-p01-dgx-c06:8000/v1,http://r01-p01-dgx-c07:8000/v1,http://r01-p01-dgx-c08:8000/v1,http://r01-p01-dgx-c09:8000/v1,http://r01-p01-dgx-c10:8000/v1,http://r01-p01-dgx-c11:8000/v1,http://r01-p01-dgx-c12:8000/v1,http://r01-p01-dgx-c13:8000/v1 --on-missing generate --on-generate delete --request-timeout 1800 --max-retries 2 --epochs 4 --gradient-accumulation-steps 8 --optimizer adamw --lr 3e-4 --weight-decay 0 --scheduler-type cosine --scheduler-warmup-ratio 0.05 --scheduler-total-steps 3605 --scheduler-min-lr-ratio 0.1 --max-grad-norm 1.0 --max-nonfinite-grad-skips 0 --fsdp-shard --checkpoint-freq 1 --checkpoint-steps 100 --retain-step-checkpoints --retain-step-checkpoints-keep-latest 4 --retain-step-checkpoints-milestone-interval 500 --log-freq 1 --logger jsonl --seed 42 --run-name muse_training2_all_global_scale16_4epochs --save-path /home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/production/checkpoints --log-dir /home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/production/logs-attempt2/train
+scripts/train.py --speculator-type dspark --block-size 7 --markov-rank 512 --markov-head-type vanilla --enable-confidence-head --confidence-head-with-markov --from-pretrained /home/inf-zixi/speculator-training/muse_spark/training2/pilots_20260909/all_global/initial --verifier-name-or-path /home/inf-zixi/.cache/huggingface/hub/models--meta-models--Muse-Spark-1.2-820B-A42B-NVFP4/snapshots/b8ace9535a34b9c33a2ef5bfbc86189f56c73e2b --target-layer-ids 14 30 46 58 62 --reuse-last-hidden-state-as-aux --draft-vocab-size 202048 --mask-token-id 200018 --draft-attn-impl simple_flex_attention --loss-implementation fused --loss-fn '{"ce": 0.1, "tv": 1.8}' --dflash-decay-gamma 4.0 --per-position-loss-weight fixed-exp-decay --max-anchors 1024 --data-path /home/inf-zixi/speculator-training/muse_spark/training2/pilots_20260909/data --train-data-ratio 0.98 --total-seq-len 32768 --noise-std 0 --num-workers 5 --prefetch-factor 2 --hidden-states-dtype bfloat16 --hidden-states-backend mooncake --mooncake-master 10.78.208.151:50051 --mooncake-metadata-server P2PHANDSHAKE --mooncake-protocol tcp --mooncake-global-segment-size 0 --mooncake-local-buffer-size 8589934592 --vllm-endpoint http://r01-p01-dgx-c01:8000/v1,http://r01-p01-dgx-c02:8000/v1,http://r01-p01-dgx-c03:8000/v1,http://r01-p01-dgx-c05:8000/v1,http://r01-p01-dgx-c06:8000/v1,http://r01-p01-dgx-c07:8000/v1,http://r01-p01-dgx-c08:8000/v1,http://r01-p01-dgx-c09:8000/v1,http://r01-p01-dgx-c10:8000/v1,http://r01-p01-dgx-c11:8000/v1,http://r01-p01-dgx-c12:8000/v1,http://r01-p01-dgx-c13:8000/v1 --on-missing generate --on-generate delete --request-timeout 1800 --max-retries 2 --epochs 4 --gradient-accumulation-steps 8 --optimizer adamw --lr 3e-4 --weight-decay 0 --scheduler-type cosine --scheduler-warmup-ratio 0.05 --scheduler-total-steps 3605 --scheduler-min-lr-ratio 0.1 --max-grad-norm 1.0 --max-nonfinite-grad-skips 0 --fsdp-shard --checkpoint-freq 1 --checkpoint-steps 100 --retain-step-checkpoints --retain-step-checkpoints-keep-latest 4 --retain-step-checkpoints-milestone-interval 500 --log-freq 1 --logger jsonl --seed 42 --run-name muse_training2_all_global_scale16_4epochs --save-path /home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/production/checkpoints --log-dir /home/inf-zixi/speculator-training/muse_spark/training2/scale_all_global_20260909/production/logs-attempt3/train