feat: Support gemma-3-1b-it

[brb-nv]

This MR adds model support for gemma-3-1b-it.

Details:

• This model has attn layers of two types - one with sliding-window attn with 10k rope base and another with regular attn with 1 million rope base.
• Most models in TRTLLM seem to be working with attention layers with attn layers of same type. Hence, there's only one set of RoPE params under AttentionParams that is passed to model's forward pass.
• However, in this case, we need two sets of RoPE params (one for each layer type mentioned above).
• Changes in this MR are adding one more set of RoPE fields under AttentionParams with a _local suffix to handle the additional layer type.
• Alternative is to copy quite some functionality from modeling_utils.py (DecoderLayerList and DecoderModelForCausalLM) and carefully orchestrate forward pass by passing a different set of AttentionParams to each layer type. I felt that's quite some code duplication and maintenance overhead.
• Sliding window requirement is mentioned to the model at runtime using max_attention_window_size with run.py or whatever API is being used.
  [512, 512, 512, 512, 512, 2048, 512, 512, 512, 512, 512, 2048, 512, 512, 512, 512, 512, 2048, 512, 512, 512, 512, 512, 2048, 512, 512]

$ rm -rf gemma3_1b_ckpt/ && python3 ./examples/gemma/convert_checkpoint.py --ckpt-type hf --model-dir ../random/hf_models/gemma-3-1b-it/ --dtype float16 --output-model-dir gemma3_1b_ckpt/
$ rm -rf gemma3_1b_eng/ && trtllm-build --checkpoint_dir gemma3_1b_ckpt/ --output_dir gemma3_1b_eng/trtllm_engine --max_batch_size 256 --max_seq_len 32768 --max_num_tokens 32768 --workers 1 --use_paged_context_fmha disable --gpus_per_node 1 --nccl_plugin auto
$ rm -rf tllm_debug/ && python3 examples/run.py --max_output_len 512 --max_input_length 2048 --input_text 'The main cities in Italy are (Write a blog post)' --engine_dir gemma3_1b_eng/trtllm_engine --tokenizer_dir ../random/hf_models/gemma-3-1b-it/ --end_id 106

[brb-nv]

/bot run

[tensorrt-cicd]

PR_Github #1179 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #1179 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #885 completed with status: 'FAILURE'

[brb-nv]

/bot run

[tensorrt-cicd]

PR_Github #1546 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #1546 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1155 completed with status: 'FAILURE'

[bebilli]

Does it support Gemma-3-27B?

[amukkara]

@brb-nv can you update examples/gemma/README.md with gemma3 instructions. specifically, clearly show instructions to set up the correct attention window values.

ideally, the attention window values should be part of the model definition so that users do not have to provide these values. we can add that in a follow-up PR.

[brb-nv]

Does it support Gemma-3-27B?

Hi, goal is to add support for text generation model first. We'll get to the multimodal models in a follow-up MR.

[brb-nv]

@brb-nv can you update examples/gemma/README.md with gemma3 instructions. specifically, clearly show instructions to set up the correct attention window values.

ideally, the attention window values should be part of the model definition so that users do not have to provide these values. we can add that in a follow-up PR.

Updated here, Anurag. 1ad80ea

[brb-nv]

/bot run

[tensorrt-cicd]

PR_Github #1646 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #1646 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1231 completed with status: 'FAILURE'

[brb-nv]

/bot run

[tensorrt-cicd]

PR_Github #1670 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #1670 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1248 completed with status: 'SUCCESS'

[chzblych]

/bot reuse-pipeline

[chzblych]

Approving this MR for merging.

[tensorrt-cicd]

PR_Github #1699 [ reuse-pipeline ] triggered by Bot

[tensorrt-cicd]

PR_Github #1699 [ reuse-pipeline ] completed with state SUCCESS
Reusing PR_Github #1670 for commit decb665