Support DeepSeekV3-style block FP8 quantization

[mgoin]

Quite a few things packed into one here, but the goal is to support the 128x128 weight and 1x128 input quantization adopted by deepseekv3 and qwen3 models. See examples: https://huggingface.co/deepseek-ai/DeepSeek-V3 and https://huggingface.co/Qwen/Qwen3-0.6B-FP8

• Added BLOCK static quantization paths for weight quantization.
• Added GROUP dynamic quantization paths for per-token-group input quantization. I feel like this is more understandable than the "1x128" block input quantization deepseek uses.
• I’ve updated all of the places where block_structure was previously treated as an “NxM” string so that it now uses a Python list of two integers (e.g. [128, 128]). I added a pydantic validator that can convert this automatically for old checkpoints that use the string.

Here is the scheme I am proposing to support this:

# Block‐wise FP8 (deepseekv3-style quantization):
# static 128x128 per‐block weights and 
# dynamic per‐token‐group activations
FP8_BLOCK = dict(
    weights=QuantizationArgs(
        num_bits=8,
        type=QuantizationType.FLOAT,
        strategy=QuantizationStrategy.BLOCK,
        symmetric=True,
        dynamic=False,
        block_structure=[128, 128],
    ),
    input_activations=QuantizationArgs(
        num_bits=8,
        type=QuantizationType.FLOAT,
        strategy=QuantizationStrategy.GROUP,
        symmetric=True,
        dynamic=True,
        observer=None,
        group_size=128,
    ),
)

[kylesayrs]

Suggested change

	block_structure: Optional[List[int]] = None
	block_structure: Optional[Tuple[int, int]] = None

[mgoin]

I feel towards keeping it a list since I don't think we can distinguish between list and tuple in the final json config

[kylesayrs]

This is mostly an aesthetic choice, but it might have aesthetic consequences if vllm wants to support fused input-weight quantization. Ex if input_quant_strategy == group and weight_quant_strategy == group

[kylesayrs]

Might want to add some validation on quant_scheme related to this as well

[dsikka]

Can you produce a test model to nm-testing andd it to this PR?

[dsikka]

Can we add a docstring explaining why this falls into the else condition?
Otherwise, I think this has grown complicated enough to easily fall prey to a bug

[kylesayrs]

@dsikka Do you have a sense of why this logic exists at all/ what cases its used? I'm not sure if inferring quant strat is really safe practice

[kylesayrs]

nit: remove

[kylesayrs]

@dsikka Do you have a sense of why this logic exists at all/ what cases its used? I'm not sure if inferring quant strat is really safe practice