DeepSeek-V2¶

DeepSeek's Mixture-of-Experts LLM, ported to pure Keras 3. Two ideas define it:

Multi-head Latent Attention (MLA): queries and key/values are compressed through low-rank bottlenecks (q_lora_rank, kv_lora_rank) before attention, which shrinks the KV cache dramatically. Each head splits into a rotary part (qk_rope_head_dim) and a non-rotary part (qk_nope_head_dim).
DeepSeekMoE: fine-grained routed experts plus always-on shared experts, with the first first_k_dense_replace layers left dense.

Memory is governed by total parameters, not active ones: every expert stays resident.

Links:

Variants¶

Load any of these with from_weights("<variant>").

Variant	Hub
`deepseek-v2-lite`	`deepseek-ai/DeepSeek-V2-Lite`
`deepseek-v2-lite-chat`	`deepseek-ai/DeepSeek-V2-Lite-Chat`
`deepseek-v2`	`deepseek-ai/DeepSeek-V2`
`deepseek-v2-chat`	`deepseek-ai/DeepSeek-V2-Chat`
`deepseek-v2.5`	`deepseek-ai/DeepSeek-V2.5`

API¶

`DeepseekV2Model`¶

The decoder backbone, no LM head. Returns {"last_hidden_state": (batch, seq, embed_dim)}.

Arg	Default	Meaning
`vocab_size`	`102400`	token vocabulary size
`embed_dim`	`2048`	model width
`num_layers`	`27`	decoder blocks
`num_heads`	`16`	query heads
`mlp_dim`	`10944`	MLP inner width
`moe_mlp_dim`	`1408`	per-expert inner width
`num_experts`	`64`	expert count
`num_experts_per_tok`	`6`	experts routed per token
`n_shared_experts`	`2`
`topk_method`	`'greedy'`
`n_group`	`1`
`topk_group`	`1`
`routed_scaling_factor`	`1.0`
`first_k_dense`	`1`
`q_lora_rank`	`None`
`kv_lora_rank`	`512`
`qk_nope_head_dim`	`128`
`qk_rope_head_dim`	`64`
`v_head_dim`	`128`
`rope_theta`	`10000.0`	rotary base frequency
`rope_scaling`	`None`
`norm_eps`	`1e-06`	RMSNorm epsilon
`max_position_embeddings`	`163840`	maximum context length
`tie_embeddings`	`False`	reuse the embedding matrix as the LM head

`DeepseekV2Generate`¶

DeepseekV2Model plus a (tied) LM head. Returns {"logits": (batch, seq, vocab_size)} and adds .generate(). Same constructor arguments as DeepseekV2Model.

generate(input_ids, attention_mask=None, max_new_tokens=None,
         eos_token_id=None, sampler=None, seed=None, **prefill_inputs)

Arg	Default	Meaning
`input_ids`	required	`(batch, seq)` token ids
`attention_mask`	`None`	`(batch, seq)` 1 = keep, 0 = padding
`max_new_tokens`	`None`	tokens to generate
`eos_token_id`	`None`	stop token (defaults to the tokenizer's)
`sampler`	`None`	sampling strategy; greedy when unset
`seed`	`None`	seed for stochastic samplers

`DeepseekV2Tokenizer`¶

Tokenizer on the tokenizers backend.

DeepseekV2Tokenizer(hf_id=None, tokenizer_file=None)

Arg	Default	Meaning
`hf_id`	`None`	Hub repo to pull the tokenizer files from
`tokenizer_file`	`None`	explicit path to a `tokenizer.json`

Calling it returns {"input_ids", "attention_mask"}, padded across the batch. It accepts a plain string or a list of strings (a batch). This tokenizer has no chat template, so pass a prompt you have already rendered rather than a message list. Decode with .decode(ids) for one sequence or .batch_decode(ids) for a batch.

End-to-end example¶

Single input¶

import os
os.environ["KERAS_BACKEND"] = "torch"   # or "jax" / "tensorflow"

from kerasformers.models.deepseek_v2 import DeepseekV2Generate, DeepseekV2Tokenizer

model = DeepseekV2Generate.from_weights("deepseek-v2-lite")
tokenizer = DeepseekV2Tokenizer.from_weights("deepseek-v2-lite")

inputs = tokenizer("Explain rotary embeddings in one sentence.")
outputs = model.generate(**inputs, max_new_tokens=64)

print(tokenizer.decode(outputs[0]))

Batch¶

Pass a list of strings. The tokenizer pads them and generate runs the batch together:

prompts = [
    "The capital of France is",
    "In one sentence, what is a transformer?",
    "Write a haiku about GPUs.",
]
inputs = tokenizer(prompts)              # {"input_ids": (3, seq), "attention_mask": (3, seq)}
outputs = model.generate(**inputs, max_new_tokens=64)

for text in tokenizer.batch_decode(outputs):
    print(text)

Backbone only¶

from kerasformers.models.deepseek_v2 import DeepseekV2Model

backbone = DeepseekV2Model.from_weights("deepseek-v2-lite")
hidden = backbone(inputs)["last_hidden_state"]   # (batch, seq, embed_dim)

Loading from the Hub¶

Any Hub repo with this architecture works via the hf: prefix, including community fine-tunes:

model = DeepseekV2Generate.from_weights("hf:deepseek-ai/DeepSeek-V2-Lite")

Lower memory¶

Larger checkpoints load in bf16 or weight-only quantized. See quantization.md:

model = DeepseekV2Generate.from_weights(
    "deepseek-v2-lite", quantization="int8", low_memory=True, load_dtype="bfloat16"
)