llmPredict built-in for LLM inference#2448
Open
kubraaksux wants to merge 8 commits intoapache:mainfrom
Open
Conversation
Register llmPredict through the full SystemDS compilation pipeline (Builtins, Opcodes, Types, DMLTranslator, HOP, LOP, CP instruction). LlmPredictCPInstruction sends HTTP POST to OpenAI-compatible servers with configurable concurrency. Includes 10 tests (7 mock, 3 live).
Introduce forward_internal with optional causal masking via upper.tri and log(1-mask) for -inf before column-wise softmax. Expose forward_causal wrapper; keep forward() signature unchanged for existing callers. Add DML + JUnit test verifying first-token invariance under causal mode when future value tokens change.
Introduces scripts/staging/llm-native/ for running pretrained HF transformer models in SystemDS. tools/convert_gpt2.py splits HF's fused c_attn projection into W_Q/W_K/W_V, upcasts to float64, and writes one CSV+MTD pair per matrix plus a manifest.json index.
Adapted from bert_layer.dml: pre-LN ordering, causal attention via multi_attention::forward_causal, and no inner final-LN. Inference- only; GELU hardcoded to gelu_new.
Pure-NumPy float64 forward over the converter's CSVs, dumping every intermediate hidden state. --compare-hf cross-checks against HF; on gpt2 (124M) every per-step max-abs-diff is below 1e-11.
Reads stacked CSV weights produced by tools/convert_gpt2.py + tools/pack_weights.py and runs a single forward pass, writing logits.csv plus per-block dumps. On gpt2 (124M), worst max-abs-diff vs the NumPy oracle is 4.5e-13, so DML logits match HF to ~1e-12. pack_weights.py exists because DML's read() requires const-string filenames; stacking per-layer matrices lets the driver row-slice inside the loop instead. Note: SystemDS/Hadoop's FileInputFormat silently skips files whose names start with '_' or '.', so input files must not use those prefixes.
compare_logits.py tokenizes a prompt, runs HF in float64, runs the NumPy oracle, optionally invokes the DML driver via subprocess, and prints a per-step max-abs-diff table across the three. Validation on gpt2 (124M): worst max|d| = 4.09e-12 at T=5 and 7.73e-12 at T=128 -- the diff stays in the float64 round-off regime and does not grow with sequence length.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Contains the
llmPredictAPI implementation (Java pipeline + tests). The full benchmark framework is in #2431. Branch also carries an end-to-end native DML inference prototype underscripts/staging/llm-native/: HF→DML weight converter, GPT-2 pre-LN block (scripts/nn/layers/gpt2_layer.dml), NumPy reference, inference driver, and a three-way correctness harness (tools/compare_logits.py). DML logits match HuggingFace within float64 round-off at T=5 and T=128. Previously tracked in #2430 (closed due to branch history issue).