Working notebook: a commonplace blog for collecting notes & exploring ideas.
Home. Site Map. Subscribe. More at expLog.

Deep Seek V3

Precis

• Showing amazing results with very little investment compared to llama models.
• Try to implement the model / run inference on it if I can.
• Lots of tricks to make the most of money invested, I’m a bit jealous

To learn

• Get better at inference
• Understand the model architecture
• Look for opportunities
• Second model to test numbers/understanding against, test simulations against (potentially try and run it locally too)

Numbers

• 14.8T tokens
• 671B params
• 37B activated per token
• 2k GPU cluster…~256 hosts?
• 14.8T tokens * (3.7 days / trillion tokens)
• 119k gpu hours for context length
• 5k gpu hours for post training
• 2.788M gpu hours total ~= 5.576M training (excluding experiments)
  • $2 per gpu hour
• 16 PP
• 64 EP spanning 8 nodes
• ZeRO-1 DP
• 61 layers
• 7168 hidden dimension
• attention head 128
• per head dimension 128
• kv compression dimension 512
• query compression dimension 1536
• all but first 3 ffn layers are moe
• each layer is 1 shared expert and 256 routed experts
• intermediate hidden dimension of expert 2048
• 8 experts per token for routed expert
• each token oges to at most 4 nodes
• prediction depth is set to 1 (one extra token)
• adamw b1 .9 b2 .95 decay .1
• start with 4k seq length
• learning rate increased slowly nd then constant, then decayed after 10t
  • annealing at the end
  • gradient clipping
  • batch size is gradually increased and then kept constant

Notes

• large moe model with 671B parameters, 37B activated per token (~5%)
• trying for performance vs cost; both efficient inference and training
• auxiliary loss free startegy for load balancing
• mullti token prediction training objective
• fp8
• dualpipe algorithm for pp (reduce bubbles)
• efficient cross node a2a kernels for ib/nvlink bandwidth
• optimized memory to avoid tensor parallelism
• 14.8T tokens
• 2 stage context length extension
  • 32k to 128k
  • SFT then RL
  • distill reasoning from deepseek R1
• chain of thought model – r1 series

architecture

• moe inside the ffn network with a router

• muti head attention inside attention block

• auxiliary load

  • measure which expert is dealing with things
  • adjust where it goes using a gating multiplier

• mtp

  • predict multiple tokens at each step
  • connect output of first token to next steps
  • speculative decoding
  • and compute a cross entropy loss as an additional training adjective
  • discarded during inference, can be repurposed to reduce latency

• infrastructure

• misc: nice visualization showing compute and communication

• HAI LLM framework

• dualpipe

• custom PTX instructions + tuning chunk size

• recompute rms norm

• ema in cpu of model params for performance estimates

  • async updates

• fp8

  • fine grained quantization
  • tile wis or block wise
  • cache and dispatch activations in fp8, optimizer states in bf16
  • validated for 1 t tokens
  • maintain precision for embedding module, output head, moe gating, normalization, attention operators
  • store master weights, weight gradients, optimizer states in higher precision

• increasing accumulation precision

• E4M3 throughout by using smaller tiles

• online quantization by deriving max values dynamically

• even further customizations

• qq: how did they choose which ones are worth handling more carefully?

• prefilling

  • redundant experts to load balance
  • based on live statistics and adjusted
  • then rearrange the gpus within a node
  • dynamic redundancy strategy

• inference / decoding

• data construction

  • multilingual, with more maths and programing
  • document packing

• tokenizer

  • byte level bpe
  • modify for optimizing compression efficiency
  • pretokenizer introduces tokens that combine punctuation and line breaks
  • randomly split further

Follow ups

• context: deep seek the company: maybe this
• r1 series of deepseek models
• read history of papers
• understand performance implications of model decisions
• low rank compression
• HAI LLM code?
• implement dualpipe
• warp specialization – ib sending, ib to nvlink, nvlink have different warps
• model performance estimates?
• follow up on gpu architecture information
• swiglu operator
• document packing

— Kunal