Add binary group quantization and WebGPU sampler fixes for Bonsai/WebLLM

[HMDRAMS-DEV]

Summary

• add q1f16_0 binary group quantization support and wire it into Qwen3 quantization registration
• fix target kind checks in WebGPU-related compiler passes so sampler/logit logic recognizes webgpu correctly
• add a regression test covering sampler attachment on WebGPU

Test plan

• run the new tests/python/compiler_pass/test_attach_sampler.py regression test
• manually verify the q1f16_0 path can be selected for Qwen3 models

Notes

• this PR contains the upstreamable mlc-llm changes only
• local Bonsai rebuild scripts and project-specific website integration were intentionally kept out of the diff

Made with Cursor

[gemini-code-assist]

Code Review

This pull request introduces binary group quantization (1-bit weights with per-group FP16 scales) to the MLC LLM framework. It includes the implementation of BinaryGroupQuantize, BinaryGroupQuantizeLinear, and BinaryGroupQuantizeEmbedding modules, along with integration into the Qwen3 model configuration. Additionally, the PR refactors target kind checks across several compiler passes to use .kind.name for consistency and adds unit tests for the GPU sampler attachment logic. Feedback is provided regarding the tensor parallel alignment check in the new quantization module, where using tirx.ceildiv is recommended to ensure consistent group count calculations when dimensions are not multiples of the group size.

[gemini-code-assist]

The tensor parallel alignment check uses in_features * num_shards // config.group_size to calculate the total number of groups. However, if the global input dimension is not a multiple of the group size, this floor division will underestimate the total group count compared to the ceildiv used for num_group on line 285. This could lead to a ValueError even if the shards are correctly sized, or conversely, allow an invalid configuration if the total groups cannot be evenly distributed. It is safer to use tirx.ceildiv for the global group count calculation to ensure consistency with the per-shard allocation.

Suggested change

	if num_shards > 1 and (in_features * num_shards // config.group_size) % num_shards != 0:
	raise ValueError(
	f"The linear dimension {in_features * num_shards} has "
	f"{in_features * num_shards // config.group_size} groups under group size "
	f"{config.group_size}. The groups cannot be evenly distributed on "
	f"{num_shards} GPUs.\n"
	"Possible solutions: reduce number of GPUs, or use quantization with smaller "
	"group size."
	)
	if num_shards > 1 and (tirx.ceildiv(in_features * num_shards, config.group_size)) % num_shards != 0:
	raise ValueError(
	f"The linear dimension {in_features * num_shards} has "
	f"{tirx.ceildiv(in_features * num_shards, config.group_size)} groups under group size "
	f"{config.group_size}. The groups cannot be evenly distributed on "
	f"{num_shards} GPUs.\n"
	"Possible solutions: reduce number of GPUs, or use quantization with smaller "
	"group size."
	)