EasyDeL / DeepSeek-V4-Flash

DeepSeek-V4-Flash (EasyDeL)

DeepSeek-V4-Flash converted to the EasyDeL checkpoint format for JAX/TPU training and serving.

Weights are bfloat16 . Nothing here is pre-quantized — quantization, if you want it, is applied at load time (see below).

Architecture

Attention is mixed per layer — sliding, compressed-sparse (rate 4) and heavily-compressed (rate 128) — with a DSA indexer selecting index_topk=512 compressed entries. MLPs are hash-routed MoE.

Loading

import easydel as ed
from jax import numpy as jnp

model = ed.AutoEasyDeLModelForCausalLM.from_pretrained(
    "EasyDeL/DeepSeek-V4-Flash",
    dtype=jnp.bfloat16,
    param_dtype=jnp.bfloat16,
    sharding_axis_dims=(1, 1, 1, 4, 1, 1),   # pp, dp, fsdp, ep, tp, sp
)

sharding_axis_dims is (pp, dp, fsdp, ep, tp, sp) . On a v5p-8, expert parallelism ( ep=4 ) is the fastest arrangement for this model; tensor parallelism works but is slower, because decode is bound by expert weight bandwidth and tp does not reduce the bytes each chip reads.

Quantized serving

model = ed.AutoEasyDeLModelForCausalLM.from_pretrained(
    "EasyDeL/DeepSeek-V4-Flash",
    dtype=jnp.bfloat16,
    param_dtype=jnp.bfloat16,
    sharding_axis_dims=(1, 1, 1, 4, 1, 1),
    quantization_config=ed.EasyDeLQuantizationConfig(
        dtype=ed.layers.quantization.QuantizationType.CHANNELWISE, bits=4
    ),
    apply_quantization=True,
)

Stacked-expert linears accept CHANNELWISE but decline block formats such as MXFP4, which would leave the MoE weights — most of the model — in bf16.

Checkpoint layout

EasyDeL/tensorstore format (zarr arrays under model/ ), not safetensors.

config.json carries fused_param_tp: 1 , meaning fused projections are stored in canonical tp=1 order; EasyDeL re-interleaves them for whatever tensor-parallel size you load at. Loading a fused quantized checkpoint at tp > 1 requires a version of EasyDeL that re-interleaves the quantized leaves — earlier versions silently produced scrambled gate/up halves at tp > 1 while being correct at tp = 1 .

Arrays are chunked at up to 8 chunks each. The upstream conversion wrote one chunk per expert (256 files per expert tensor, ~35k files total), which exceeds the Hub's 20,000-file-per-repo limit; the content is identical, only the chunk grid differs.

License

Follows the license of the original DeepSeek-V4 release.