release/v1.20.0

Release Summary

New Features

• SpD, multiprojection heads (#306): Implemented post-attention hidden size projections to speculate tokens ahead of the base model.
• Adding compilation support for io_encrypt flag (#393): Added support for Model-IP I/O encryption feature using qaic-exec (compile only).
• Multi model end to end test pipelines (#337): Introduced comprehensive test pipelines for multi-model scenarios.
• QNN Compilation path Support in QEFFBaseModel class (#374): Added support for QNN compilation path in the QEFFBaseModel class.
• CLI support for multimodel (AutoModelForImageTexttoText) (#287): Enabled CLI support for multimodel .
• Support for Disaggregated serving (#365): Added support for disaggregated serving.
• Support for GGUF model execution (without quantized weights) (#368): Enabled GGUF model execution without quantized weights.
• SwiftKV backup Pending changes (#367): Implemented pending changes for SwiftKV backup.
• Upgrading transformers version to latest 4.50 (#331): Upgraded transformers to version 4.50.
• Added support for gradient checkpointing in the finetuning script (#338): Enabled gradient checkpointing in the finetuning script.
• Passing device type in torch GradScaler (#345): Added functionality to pass device type in torch GradScaler.
• Enabled FP8 models for replicate_kv_heads script (#353): Enabled FP8 models for the replicate_kv_heads script.
• QNN Compilation command changes (#327): Updated QNN compilation commands.

Newly Onboarded Models

• Granite Vision (#359): Onboarded Granite Vision model.
• Granite MOE (#362): Onboarded Granite MOE model.

Bug Fixes

• Config dump issue when apps sdk not installed for qnn execution (#379): Fixed config dump issue when apps SDK is not installed for QNN execution.
• Removed device IDs from the test (#389): Removed device IDs from the test suite.
• Compilation issue of VLMs with mxint8(kv) precision (#336): Fixed compilation issue of VLMs with mxint8(kv) precision.
• Retrying downloads logic for stability (#370): Improved stability by updating retry logic for downloads.
• Mllama bug fix (#369): Fixed bug in Mllama.
• Update qeff wheel package name (#352): Updated the qeff wheel package name.
• Fix replicates biases too if they exist (e.g., Qwen) (#328): Fixed replication biases if they exist (e.g., Qwen).

Release V1.19.3

Added Features

• Vision Language Model
• Speech Sequence to Sequence Model
• Support for FP8 Execution
• Prompt-Lookup Decoding sample script.