GitLab

Deepseek V2 is a MoE model from Deepseek. Relevant variations
compared to other models:

- Grouped top-K in expert selection.
- mscale in yarn is calculated using the `mscale` and `mscale_all_dim`
  configuration options.
- `mscale_all_dim` is also used in scaling attention softmax.
- Permuting of the query/key representations before applying rotary
  embeddings.
- Some projections cannot be sharded (`q_a_proj`, `kv_a_proj_with_mqa`).
  So, we need weight loads that supports quantized weights. To this
  end `{Weights,WeightLoader}.get_weight` was added.
- The query/key head dimensionality differs from that of the value,
  so we need to pad during attention.
- Heads with size 192, needs an extension to our paged attention
  fork and we need to ensure that the KV cache is allocated with the
  correct size.
- Shared experts.

* Improve the handling of quantized weights

Handling of quantized weights was split between two mechanisms:

- For quantized checkpoints, we used the new weight loader
  infrastructure.
- For quantization while loading (EETQ, FP8, bitsandbytes) we
  instead relied on conditional in `get_linear`.

Weight loaders support context managers to selectively load
particular layers with different weight loaders, which is useful
for models like Idefics2 AWQ, which uses a quantized text model,
but unquantized vision and connector models. However, the context
manager would be overrided by `get_linear`, which string-checks
`quantizer`. Also, the context manager would not work with
EETQ, FP8, and bitsandbytes.

This change migrates all quantizers to the weight loader infrastructure.
This has several benefits:

- We can use context managers with all quantizers.
- All the implementation details move down to the quantizer layers,
  `get_linear` does not need to know how to handle quantizer linear
  layers.
- All quantizer weights are strongly typed, we don't pass around
  raw tensors.
- We don't have to pass around the `quantizer` string everywhere.

* Exclude non-MLP layers when using FP8 quantization with Llama

Quantized weights were loaded in the `Weights` class, but this was
getting quite unwieldy, where every higher level method to load weights
was a long conditional to cover all the different quantizers.

This change moves loading of quantized weights out of the `Weights`
class. This is done by defining a simple `WeightsLoader` interface
that is implemented by `Exl2WeightsLoader`, `GPTQWeightsLoader`,
and `MarlinWeightsLoader`. These implementations are in the quantizers'
respective modules. The `Weights` class provides the low-level load
operations (such as loading tensors or sharded tensors), but delegates
loads that need quantizer-specific weight processing to a loader. The
loaders still use the low-level functionality provided by `Weights`.

I initially tried making a hierarchy where a class like `GPTQWeights`
would inherit from `Weights`. But it is not very flexible (e.g. does
not work well with the new weight storage mock used in tests) and
the implicit indirections made the code harder to follow.

Mostly straightforward, changes to existing code:

* Wrap quantizer parameters in a small wrapper to avoid passing
  around untyped tuples and needing to repack them as a dict.
* Move scratch space computation to warmup, because we need the
  maximum input sequence length to avoid allocating huge
  scratch buffers that OOM.