Auto max prefill (#2797)

* Attempt at automatic max batch prefill. * Taking into account number of shards. * Adding more cards. * Adding A100 + H100 * Adding a few more cards. * Logprobs cost too much. * h100 better name, and keep factor of 2 * Damn inflated sparse tflops. * Typo in h100. * Updated the flops calculation (checked with fvcore). * chunking by default. * Fix prefix caching for chat completion since we removed logprobs. * More tests. * Dropping all the prefill logprobs. * Add a flag that enables users to get logprobs back. * Repairing prompt token counting. * Fixing a few tests. * Remove some scaffolding. * Attempting to reduces the issues (workarounds for now).

Auto max prefill (#2797)
* Attempt at automatic max batch prefill. * Taking into account number of shards. * Adding more cards. * Adding A100 + H100 * Adding a few more cards. * Logprobs cost too much. * h100 better name, and keep factor of 2 * Damn inflated sparse tflops. * Typo in h100. * Updated the flops calculation (checked with fvcore). * chunking by default. * Fix prefix caching for chat completion since we removed logprobs. * More tests. * Dropping all the prefill logprobs. * Add a flag that enables users to get logprobs back. * Repairing prompt token counting. * Fixing a few tests. * Remove some scaffolding. * Attempting to reduces the issues (workarounds for now).
5df80590 · Nicolas Patry · GitHub · 8c3669b2 · 5df80590 · 5df80590
Unverified Commit 5df80590 authored Dec 06, 2024 by Nicolas Patry Committed by GitHub Dec 06, 2024
10 changed files
--- a/load_tests/long.js
+++ b/load_tests/long.js
--- a/load_tests/long.py
+++ b/load_tests/long.py
--- a/load_tests/long_prompt2.py
+++ b/load_tests/long_prompt2.py
--- a/load_tests/orca.py
+++ b/load_tests/orca.py
--- a/router/src/config.rs
+++ b/router/src/config.rs
--- a/router/src/lib.rs
+++ b/router/src/lib.rs
--- a/router/src/server.rs
+++ b/router/src/server.rs
--- a/router/src/vertex.rs
+++ b/router/src/vertex.rs
--- a/server/text_generation_server/models/flash_causal_lm.py
+++ b/server/text_generation_server/models/flash_causal_lm.py
--- a/server/text_generation_server/models/globals.py
+++ b/server/text_generation_server/models/globals.py