[None][chore] Remove legacy TensorRT examples, docs and tests

docs/source/blogs/Best_perf_practice_on_DeepSeek-R1_in_TensorRT-LLM.md

@@ -414,13 +414,13 @@ Note that, `max_batch_size` and `max_num_tokens` can easily affect the performan
 
 Generally, you should make sure that `max_batch_size` is not too low to bottleneck the throughput, and `max_num_tokens` needs to be large enough so that it covers the max input sequence length of the samples in dataset, as mentioned in below section "WIP: Chunked context support on DeepSeek models".
 
-For more details on `max_batch_size` and `max_num_tokens`, refer to [Tuning Max Batch Size and Max Num Tokens](../legacy/performance/performance-tuning-guide/tuning-max-batch-size-and-max-num-tokens.md).
+For more details on `max_batch_size` and `max_num_tokens`, refer to the performance tuning guidance in the developer documentation.
 
 ### MLA chunked context
 
 MLA currently supports the chunked context feature on both Hopper and Blackwell GPUs. You can use `--enable_chunked_context` to enable it. This feature is primarily designed to reduce TPOT (Time Per Output Token). The default chunk size is set to `max_num_tokens`. If you want to achieve a lower TPOT, you can appropriately reduce the chunk size. However, please note that this will also decrease overall throughput. Therefore, a trade-off needs to be considered.
 
-For more details on `max_num_tokens`, refer to [Tuning Max Batch Size and Max Num Tokens](../legacy/performance/performance-tuning-guide/tuning-max-batch-size-and-max-num-tokens.md).
+For more details on `max_num_tokens`, refer to the performance tuning guidance in the developer documentation.
 
 ### Out of memory issues
 

docs/source/blogs/XQA-kernel.md

@@ -2,7 +2,7 @@
 
 XQA kernel provides optimization for [MQA](https://arxiv.org/abs/1911.02150) and [GQA](https://arxiv.org/abs/2305.13245v3) during generation phase. It also provides optimization for beam search. Using tensor cores for acceleration, reducing data loading and conversion, it delivers increased throughput within the same latency budget. Increased throughput allows serving greater number of user requests while providing the same experience.
 
-Support matrix and usage flags are described in [docs/source/legacy/advanced/gpt-attention](../legacy/advanced/gpt-attention.md#xqa-optimization).
+Support matrix and usage flags are described in the [Attention](../features/attention.md) documentation.
 
 **Increased Throughput:**
 Looking at the Throughput-Latency curves below, we see that the enabling of XQA optimization increases throughput. Higher throughput equates to serving more users, and we can see that TPOT on the Y-axis flattens out when XQA gets enabled.

docs/source/legacy/advanced/lowprecision-pcie-allreduce.md → docs/source/features/lowprecision-allreduce.md

@@ -5,8 +5,7 @@ Note:
 This feature is optimized for PCIe-based GPU topologies and may affect model accuracy. Please evaluate precision impact for your specific workload.
 ```
 
-
-TRT-LLM supports `low-precision-allreduce`, a communication optimization that accelerates AllReduce operations in PCIe-based GPU environments. This feature quantizes FP16/BF16 data to FP8 during network transmission, reducing communication volume and improving performance.
+TensorRT LLM supports `low-precision-allreduce`, a communication optimization that accelerates AllReduce operations in PCIe-based GPU environments. This feature quantizes FP16/BF16 data to FP8 during network transmission, reducing communication volume and improving performance. It is available on the PyTorch backend through the `allreduce_strategy` field of `LlmArgs`.
 
 ## Algorithm
 
@@ -35,20 +34,16 @@ Low-Precision-AllReduce is specifically designed for the topology shown above, w
 
 ## Usage
 
-The Low-Precision-AllReduce algorithm can be enabled in two ways:
+Enable the Low-Precision-AllReduce algorithm by setting the `allreduce_strategy` field in `LlmArgs`:
 
-1. **Direct specification** in your code:
-```
-AllReduce allreduce(mapping=mapping, strategy=AllReduceStrategy.LOWPRECISION);
-```
+```python
+from tensorrt_llm import LLM
 
-2. Enable by LlmArgs
-```
-Set allreduce_strategy field in LlmArgs.
-Candidates of strategies are "AUTO", "NCCL", "UB", "MINLATENCY", "ONESHOT", "TWOSHOT", "LOWPRECISION" and "MNNVL".
-If no strategy is set, AUTO will be set.
+llm = LLM(model="<hf_model>", allreduce_strategy="LOWPRECISION")
 ```
 
+The candidate strategies are `"AUTO"`, `"NCCL"`, `"UB"`, `"MINLATENCY"`, `"ONESHOT"`, `"TWOSHOT"`, `"LOWPRECISION"` and `"MNNVL"`. If no strategy is set, `AUTO` is used.
+
 ## Performance and Accuracy Considerations
 
 Low-Precision-AllReduce reduces communication volume by using FP8 data format for transmission. This optimization:
@@ -58,4 +53,4 @@ Low-Precision-AllReduce reduces communication volume by using FP8 data format fo
 
 Users should evaluate the precision impact on their specific models and workloads.
 
-**Note**: When compiling TensorRT-LLM without enabling the `ENABLE_FP8` option, setting Low Precision allreduce will not take effect.
+**Note**: When TensorRT LLM is built without the `ENABLE_FP8` option, setting the low-precision AllReduce strategy will not take effect.

docs/source/index.rst

@@ -72,6 +72,7 @@ Welcome to TensorRT LLM's Documentation!
    features/overlap-scheduler.md
    features/paged-attention-ifb-scheduler.md
    features/parallel-strategy.md
+   features/lowprecision-allreduce.md
    features/quantization.md
    features/sampling.md
    features/additional-outputs.md
@@ -127,10 +128,9 @@ Welcome to TensorRT LLM's Documentation!
 
 .. toctree::
    :maxdepth: 2
-   :caption: Use TensorRT Engine
-   :hidden:
+   :caption: Migration
 
-   legacy/tensorrt_quickstart.md
+   legacy/tensorrt-backend-removal.md
 
 Indices and tables
 ==================