Chunk prefill cache writes, remove div_i32 from insert_or_update_cache #289
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
vllm/hpu/utils.py
Outdated
| num_slots_available, block_indices, | ||
| block_offset) | ||
| def forward(self, input, cache, block_indices, block_offset): | ||
| insert_or_update_cache(input, cache, block_indices, block_offset) |
There was a problem hiding this comment.
Isn't VLLMKVCache patched by INC/HQT? Will this work with fp8?
There was a problem hiding this comment.
This actually breaks FP8 ([rank0]: TypeError: PatchedVLLMKVCache.forward() missing 2 required positional arguments: 'block_indices' and 'block_offset').
madamczykhabana
approved these changes
Sep 26, 2024
michalkuligowski
deleted the
private/kzawora/insert_or_update_cache_opt
branch
zhouyu5
pushed a commit
to zhouyu5/vllm-fork
that referenced
this pull request
Sep 27, 2024
HabanaAI#289) Re-implements following PRs for current habana_main: HabanaAI#102 (Removing div_i32 operations from each layer) HabanaAI#115 (removing scatter for reshape&cache in case of prompt) Accuracy (GSM8K on Llama3.1-8B-Instruct): | Tasks |Version| Filter |n-shot| Metric | |Value | |Stderr| |---------------|------:|----------------|-----:|-----------|---|-----:|---|-----:| |gsm8k_cot_llama| 3|flexible-extract| 8|exact_match|↑ |0.8415|± |0.0101| | | |strict-match | 8|exact_match|↑ |0.8400|± |0.0101| I've benchmarked this change on Llama3.1-8B-Instruct and on average, +2.50% throughput gain (+558.14 tok/s, ~21594 tok/s -> ~22152 tok/s) can be observed across all prefill buckets on G2, with up to +4.40% (+956.79 tok/s, ~25031 -> ~25988 tok/s) throughput increase in compute-bound scenarios.
zhouyu5
pushed a commit
to zhouyu5/vllm-fork
that referenced
this pull request
Sep 27, 2024
HabanaAI#289) Re-implements following PRs for current habana_main: HabanaAI#102 (Removing div_i32 operations from each layer) HabanaAI#115 (removing scatter for reshape&cache in case of prompt) Accuracy (GSM8K on Llama3.1-8B-Instruct): | Tasks |Version| Filter |n-shot| Metric | |Value | |Stderr| |---------------|------:|----------------|-----:|-----------|---|-----:|---|-----:| |gsm8k_cot_llama| 3|flexible-extract| 8|exact_match|↑ |0.8415|± |0.0101| | | |strict-match | 8|exact_match|↑ |0.8400|± |0.0101| I've benchmarked this change on Llama3.1-8B-Instruct and on average, +2.50% throughput gain (+558.14 tok/s, ~21594 tok/s -> ~22152 tok/s) can be observed across all prefill buckets on G2, with up to +4.40% (+956.79 tok/s, ~25031 -> ~25988 tok/s) throughput increase in compute-bound scenarios.
This was referenced Nov 12, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Re-implements following PRs for current habana_main:
#102 (Removing div_i32 operations from each layer)
#115 (removing scatter for reshape&cache in case of prompt)
Accuracy (GSM8K on Llama3.1-8B-Instruct):
I've benchmarked this change on Llama3.1-8B-Instruct and on average, +2.50% throughput gain (+558.14 tok/s, ~21594 tok/s -> ~22152 tok/s) can be observed across all prefill buckets on G2, with up to +4.40% (+956.79 tok/s, ~25031 -> ~25988 tok/s) throughput increase in compute-bound scenarios.