schema: dgx.deepseek-v4-vision-nvfp4.v1

source:
  repository: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
  revision: 86f746b36186f0e567729a5c06a8c918caba82a9
  license: MIT
  expected_shards: 48
  expected_tensor_bytes: 167811372792
  local_path: /var/tmp/models/deepseek-v4-flash-vision-exp-86f746b3

runtime:
  interconnect: QFSP
  interconnect_interface: enp1s0f0np0
  nodes:
    - ssh_target: spark-1.local
      rank: 0
      cluster_address: 192.168.100.10
    - ssh_target: spark-2.local
      rank: 1
      cluster_address: 192.168.100.11
  model_parallel: 2
  container: qwen38-flash-next-nvfp4-runtime:sm121-skinny-bf16-v2
  modelopt_repository: NVIDIA/Model-Optimizer
  modelopt_revision: 029c67f27e67088fb19ac0a9af241dc2bc740650

paths:
  mp_checkpoint: /var/tmp/models/deepseek-v4-flash-vision-exp-mp2-mxfp4
  amax: /var/tmp/models/deepseek-v4-flash-vision-exp-nvfp4-amax
  output: /var/tmp/models/deepseek-v4-flash-vision-exp-nvfp4

quantization:
  algorithm: NVFP4_W4A4
  scope: routed_experts_only
  group_size: 16
  source_expert_format: MXFP4_E2M1_UE8M0_block32
  weight_conversion: lossless_mxfp4_to_nvfp4
  cast_flag: --cast_mxfp4_to_nvfp4
  preserved:
    - attention
    - router_gate
    - shared_experts
    - embeddings
    - lm_head
    - mtp
    - vision_tower

calibration:
  datasets:
    - cnn_dailymail
    - nemotron-post-training-dataset-v2
  samples_per_dataset: 64
  sequence_length: 512
  batch_size_per_rank: 1
  tokenizer_padding_side: left
  checkpoint_load: tensor_streaming_batched_mmap_with_page_cache_eviction
  checkpoint_mmap_chunk_tensors: 128
  memory_watchdog_minimum_available_gib: 12
  memory_watchdog_consecutive_samples: 2
  memory_watchdog_emergency_available_gib: 8
  sm121_sparse_attention:
    scope: calibration_only
    source_block: 64
    calibration_block: 32
    source_dynamic_shared_memory_bytes: 104448
    gb10_optin_shared_memory_bytes: 101376
  hadamard_transform:
    scope: calibration_only
    source_dependency: fast_hadamard_transform_unversioned
    implementation: pytorch_bf16_butterfly
    dimension: 128
  statistic: max

validation:
  require_lossless_block_ratio: 1.0
  require_exact_key_transform: true
  require_passthrough_key_placement: true
  require_finite_positive_scales: true

publication:
  repository: s-zaizen/DeepSeek-V4-Flash-Vision-Exp-NVFP4
  visibility: public