docs: Update LLM getting-started guide to PyTorch backend

[faradawn]

Deprecating the engine build in favor of pytorch / llmapi backend.

Triton Latest release: 26.07. TRTLLM container 1.2.1. Thus using DSR1 as example. DSV4 and GLM-5.2 are not supported in this TRTLLM yet.

Tested on B200x8.

[greptile-apps]

Greptile Summary

This PR replaces the deprecated Phi-3 TRT-LLM engine-build guide (~1,700 lines) with a concise (~185-line) getting-started guide for the modern LLM API / PyTorch backend, using nvidia/DeepSeek-R1-0528-FP4-V2 as the example model.

• Workflow modernized: removes the convert_checkpoint.py + trtllm-build + inflight_batcher_llm ensemble path and replaces it with a single model.yaml-driven LLM API workflow; the git clone is now pinned to a matching TRT-LLM version tag.
• Hardware scope narrowed: the chosen example model (FP4-V2) requires Blackwell (B200+) GPUs; this is only mentioned in a resource-usage note in step 4, after a user would already have configured and potentially started downloading the 800 GB+ model.
• max_batch_size: 0 note is misleading: the benchmark section states this setting means "no batching", but for the LLM API backend it instead defers batching to the TRT-LLM continuous-batching scheduler.

Confidence Score: 4/5

• The rewrite is a clear improvement overall, but the primary example model requires Blackwell (B200+) GPUs due to FP4 quantization — a requirement that is not surfaced until well into the guide, after a reader may have already started configuring or downloading the model.
• The guide correctly pins the TRT-LLM git tag to the container's library version and simplifies the workflow significantly. The one substantive issue is that the hardware prerequisite (Blackwell GPU required for FP4 inference) is buried in a resource-usage note rather than stated upfront, meaning readers on A100 or H100 hardware will not know they're on an unsupported path until the server fails to start.
• docs/getting_started/llm.md — the hardware prerequisite section and the benchmark batching note both need attention before this guide will be accurate for all readers.

Important Files Changed

Filename

Overview

docs/getting_started/llm.md

Complete rewrite (~1,700 → ~185 lines) replacing the deprecated Phi-3 TRT-LLM engine-build workflow with the modern LLM API / PyTorch backend using DeepSeek-R1-0528-FP4-V2. The new guide is much cleaner and uses a pinned git tag. One real usability issue: the FP4 model's Blackwell-only hardware requirement is not surfaced prominently enough for readers who may download the model before discovering the incompatibility.

Sequence Diagram

sequenceDiagram
    participant User
    participant Docker as Triton Container
    participant HF as Hugging Face Hub
    participant TRTLLM as TRT-LLM Repo v1.2.1
    participant Triton as Triton Server

    User->>Docker: docker run step 1
    User->>Docker: check tensorrt_llm version
    Docker-->>User: e.g. 1.2.1
    User->>TRTLLM: git clone depth 1 branch v1.2.1 step 2
    User->>TRTLLM: Edit model.yaml step 3
    User->>Docker: python3 launch_triton_server.py step 4
    Docker->>HF: Pull model weights
    HF-->>Docker: Model weights
    Docker->>Triton: Start with LLM API backend
    Triton-->>User: HTTP 8000 gRPC 8001 Metrics 8002
    User->>Triton: POST generate
    Triton-->>User: Completion response
    User->>Triton: POST generate_stream with decoupled True
    Triton-->>User: Streamed tokens

Loading

_{Reviews (9): Last reviewed commit: "docs: Fix two broken commands found by v..." | Re-trigger Greptile}

[whoisj]

When do these changes take effect? If they already apply, then I can approve now. If Triton needs to update to a newer version of TRTLLM, then we should wait until then.