Skip to content

Mistral batched decode: LLaMA decode path + parameterized RoPE theta (12x batched, full serving stack) #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mikepapadim wants to merge 1 commit into
base: feat/static-batched-decode
Choose a base branch
from
Open

Mistral batched decode: LLaMA decode path + parameterized RoPE theta (12x batched, full serving stack) #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions BATCHED_DECODE.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,15 @@ capped lower):
| Llama-3.2-1B | 16 · 2048 · 128256 | 101 tok/s | 128 | 4175 tok/s | 41× |
| Qwen3-1.7B | 28 · 2048 · 151936 | 48 tok/s | 64 | 1433 tok/s | 30× |
| Qwen3-4B | 36 · 2560 · 151936 | 39 tok/s | 32 | 405 tok/s | 10× |
| Mistral-7B-v0.3 | 32 · 4096 · 32768 | 27 tok/s | 32 | 331 tok/s | 12× |

**Supported models: LLaMA, Qwen3, Mistral (FP16).** Mistral runs on the LLaMA decode path with
no new kernels — the only generalizations were typing the layer graph to the base
`Configuration` and parameterizing the RoPE theta (`config.ropeTheta()`, e.g. Mistral 1e6). Any
LLaMA-family model (RMSNorm + SwiGLU + GQA + RoPE, no QK-norm, no sliding window) drops in the
same way; the full serving stack (continuous + paging + prefix + on-device sampling) is
model-agnostic and applies unchanged (Mistral-7B: 128-request continuous+paged+prefix →
prefix-consistent, ~11.6× less KV, 82.8% fewer prefill tokens).

All bit-exact vs the single-stream greedy reference (`all B streams identical: true`)
and coherent.
Expand Down
4 changes: 2 additions & 2 deletions src/main/java/org/beehive/gpullama3/bench/BatchedDecodeEngine.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,10 +158,10 @@ public static void main(String[] args) throws Exception {
} else {
LlamaFP16LayersBatchDecodeMMA l = paged
? new LlamaFP16LayersBatchDecodeMMA((LlamaState) state, (LlamaTornadoWeights) weights,
(LlamaConfiguration) config, B, decodeCtx, keyCacheBatch, valueCacheBatch, seqPositions,
config, B, decodeCtx, keyCacheBatch, valueCacheBatch, seqPositions,
blockTable, blockSize, maxBlocksPerSlot)
: new LlamaFP16LayersBatchDecodeMMA((LlamaState) state, (LlamaTornadoWeights) weights,
(LlamaConfiguration) config, B, decodeCtx, keyCacheBatch, valueCacheBatch, seqPositions);
config, B, decodeCtx, keyCacheBatch, valueCacheBatch, seqPositions);
layerITGs = l.getLayerImmutableTaskGraphs();
lastLayerId = l.getLastLayerTaskGraphID();
updateLayerSched = l::updateGridScheduler;
Expand Down
8 changes: 4 additions & 4 deletions src/main/java/org/beehive/gpullama3/tornadovm/kernels/TransformerBatchPrefillKernels.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1658,7 +1658,7 @@ public static void batchedDecodeRopeWithKVCachePacked(KernelContext context,
FloatArray wrapValueCache,
int kvDim, int headSize,
int layerIndex, int numLayers,
int contextLength, int dim) {
int contextLength, int dim, float ropeTheta) {
int globalIdx = context.globalIdx;
int halfDim = dim / 2;
int batchIdx = globalIdx / halfDim;
Expand All @@ -1673,7 +1673,7 @@ public static void batchedDecodeRopeWithKVCachePacked(KernelContext context,

if (i + 1 < dim) {
int head_dim = i % headSize;
float freq = 1.0f / TornadoMath.pow(50000.0f, head_dim / (float) headSize);
float freq = 1.0f / TornadoMath.pow(ropeTheta, head_dim / (float) headSize);
float val = pos * freq;
float fcr = TornadoMath.cos(val);
float fci = TornadoMath.sin(val);
Expand Down Expand Up @@ -1830,7 +1830,7 @@ public static void batchedDecodePagedRopeWithKVCachePacked(KernelContext context
FloatArray valuePool,
int kvDim, int headSize,
int layerIndex, int numLayers,
int blockCfg, int dim) {
int blockCfg, int dim, float ropeTheta) {
int blockSize = blockCfg & 0xFFFF;
int maxBlocksPerSlot = blockCfg >>> 16;
int globalIdx = context.globalIdx;
Expand All @@ -1847,7 +1847,7 @@ public static void batchedDecodePagedRopeWithKVCachePacked(KernelContext context

if (i + 1 < dim) {
int head_dim = i % headSize;
float freq = 1.0f / TornadoMath.pow(50000.0f, head_dim / (float) headSize);
float freq = 1.0f / TornadoMath.pow(ropeTheta, head_dim / (float) headSize);
float val = pos * freq;
float fcr = TornadoMath.cos(val);
float fci = TornadoMath.sin(val);
Expand Down
12 changes: 6 additions & 6 deletions ...rg/beehive/gpullama3/tornadovm/layers/type/fp16/decode/LlamaFP16LayersBatchDecodeMMA.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@

import org.beehive.gpullama3.inference.state.LlamaState;
import org.beehive.gpullama3.inference.weights.tornado.LlamaTornadoWeights;
import org.beehive.gpullama3.model.llama.LlamaConfiguration;
import org.beehive.gpullama3.model.Configuration;
import org.beehive.gpullama3.tornadovm.kernels.TransformerBatchPrefillKernels;
import org.beehive.gpullama3.tornadovm.scheduling.WorkerGridFactory;
import uk.ac.manchester.tornado.api.GridScheduler;
Expand Down Expand Up @@ -42,7 +42,7 @@ public class LlamaFP16LayersBatchDecodeMMA {

private final LlamaState state;
private final LlamaTornadoWeights weights;
private final LlamaConfiguration config;
private final Configuration config;
private final KernelContext context = new KernelContext();
private final int batchSize;
private final int paddedBatch;
Expand All @@ -59,7 +59,7 @@ public class LlamaFP16LayersBatchDecodeMMA {
private String lastLayerTaskGraphID;

public LlamaFP16LayersBatchDecodeMMA(LlamaState state, LlamaTornadoWeights weights,
LlamaConfiguration config, int batchSize, int decodeCtx,
Configuration config, int batchSize, int decodeCtx,
FloatArray keyCacheBatch, FloatArray valueCacheBatch,
IntArray seqPositions) {
this(state, weights, config, batchSize, decodeCtx, keyCacheBatch, valueCacheBatch, seqPositions,
Expand All @@ -68,7 +68,7 @@ public LlamaFP16LayersBatchDecodeMMA(LlamaState state, LlamaTornadoWeights weigh

/** Paged constructor: keyCacheBatch/valueCacheBatch are the shared block pools. */
public LlamaFP16LayersBatchDecodeMMA(LlamaState state, LlamaTornadoWeights weights,
LlamaConfiguration config, int batchSize, int decodeCtx,
Configuration config, int batchSize, int decodeCtx,
FloatArray keyCacheBatch, FloatArray valueCacheBatch,
IntArray seqPositions,
IntArray blockTable, int blockSize, int maxBlocksPerSlot) {
Expand Down Expand Up @@ -172,7 +172,7 @@ private TaskGraph createLayerTaskGraph(int layerIndex) {
context, seqPositions, blockTable,
state.qkvResultBatch, keyCacheBatch, valueCacheBatch,
kvDim, config.headSize(), layerIndex, config.numberOfLayers(),
blockCfg, dim);
blockCfg, dim, config.ropeTheta());

g.task("batch_attention",
TransformerBatchPrefillKernels::batchedDecodePagedAttentionFP16Out,
Expand All @@ -188,7 +188,7 @@ private TaskGraph createLayerTaskGraph(int layerIndex) {
context, seqPositions,
state.qkvResultBatch,
keyCacheBatch, valueCacheBatch,
kvDim, config.headSize(), layerIndex, config.numberOfLayers(), decodeCtx, dim);
kvDim, config.headSize(), layerIndex, config.numberOfLayers(), decodeCtx, dim, config.ropeTheta());

g.task("batch_attention",
TransformerBatchPrefillKernels::batchedDecodeAttentionFP16Out,
Expand Down