Skip to content

Actions: google/gemma.cpp

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
2,500+ workflow runs
2,500+ workflow runs

Filter by Workflow

⌥ + click/return to exclude

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
Merge pull request #1026 from Magnushst:fix/flash-attention-nf build #8133: Commit b68def1 pushed by copybara-service Bot
27m 0s dev
dev
27m 0s
Avoids repeated heap allocations/deallocations of `layout.total_bytes… build #8132: Commit ffc1abc pushed by copybara-service Bot
26m 3s dev
dev
26m 3s
Avoids repeated heap allocations/deallocations of layout.total_bytes per attention subtask and kv_out_mem per token step by storing worker_workspaces and kv_out_mem on AttentionActivations / AttentionActivationsPtrs. This eliminates ~20M minor page faults and improves decode throughput by +18% to +26% on AMD Turin. build #8131: Pull request #1025 synchronize by copybara-service Bot
26m 28s test_979178228
test_979178228
26m 28s
Avoids repeated heap allocations/deallocations of `layout.total_bytes… build #8130: Commit ffc1abc pushed by copybara-service Bot
26m 40s test_979178228
test_979178228
26m 40s
Avoids repeated heap allocations/deallocations of layout.total_bytes per attention subtask and kv_out_mem per token step by storing worker_workspaces and kv_out_mem on AttentionActivations / AttentionActivationsPtrs. This eliminates ~20M minor page faults and improves decode throughput by +18% to +26% on AMD Turin. build #8129: Pull request #1025 synchronize by copybara-service Bot
26m 36s test_979178228
test_979178228
26m 36s
Avoids repeated heap allocations/deallocations of `layout.total_bytes… build #8128: Commit f951eff pushed by copybara-service Bot
27m 13s test_979178228
test_979178228
27m 13s
Replaces the second ParallelFor barrier/reduction loop in LocalAttentionForAllHeadsTokensAndBatch (tiled_attention.cc) with a lock-free 64-bit atomic compare-and-exchange state machine (AtomicGroupMergeSlot) partitioned into groups of up to 32 queries (kMergeGroupSize = 32). Subtasks hand off completed 32-query group pointers immediately via CAS or claim and merge waiting groups outside any lock. Uses Highway SIMD vectorization (MergeOnlineSoftmaxGroup32, hn::FastExpMinusOrZero) to compute online softmax exponents and scaling factors across all 32 queries simultaneously and combine attention outputs in a single FMA pass. build #8127: Pull request #1028 opened by copybara-service Bot
25m 30s test_979178229
test_979178229
25m 30s
Replaces the second ParallelFor barrier/reduction loop in `LocalAtt… build #8126: Commit 1956586 pushed by copybara-service Bot
26m 14s test_979178229
test_979178229
26m 14s
Avoids repeated heap allocations/deallocations of layout.total_bytes per attention subtask and kv_out_mem per token step by storing worker_workspaces and kv_out_mem on AttentionActivations / AttentionActivationsPtrs. This eliminates ~20M minor page faults and improves decode throughput by +18% to +26% on AMD Turin. build #8125: Pull request #1025 synchronize by copybara-service Bot
26m 7s test_979178228
test_979178228
26m 7s
Avoids repeated heap allocations/deallocations of `layout.total_bytes… build #8124: Commit 4535389 pushed by copybara-service Bot
23m 15s test_979178228
test_979178228
23m 15s
Use the current SIMD target's width for the transposed KV cache build #8123: Pull request #1026 labeled by jan-wassenberg
20m 54s Magnushst:fix/flash-attention-nf
Magnushst:fix/flash-attention-nf
20m 54s
Use the current SIMD target's width for the transposed KV cache build #8122: Pull request #1026 synchronize by Magnushst
25m 25s Magnushst:fix/flash-attention-nf
Magnushst:fix/flash-attention-nf
25m 25s
Use the current SIMD target's width for the transposed KV cache build #8121: Pull request #1027 opened by Magnushst
40s Magnushst:fix/flash-attention-nf
Magnushst:fix/flash-attention-nf
40s
Use the current SIMD target's width for the transposed KV cache GitHub Actions Scan #9: Pull request #1027 opened by Magnushst
1m 7s
1m 7s
Use the current SIMD target's width for the transposed KV cache build #8120: Pull request #1026 opened by Magnushst
1m 13s Magnushst:fix/flash-attention-nf
Magnushst:fix/flash-attention-nf
1m 13s
Avoids repeated heap allocations/deallocations of layout.total_bytes per attention subtask and kv_out_mem per token step by storing worker_workspaces and kv_out_mem on AttentionActivations / AttentionActivationsPtrs. This eliminates ~20M minor page faults and improves decode throughput by +18% to +26% on AMD Turin. build #8119: Pull request #1025 opened by copybara-service Bot
26m 9s test_979178228
test_979178228
26m 9s
Avoids repeated heap allocations/deallocations of `layout.total_bytes… build #8118: Commit 5c1be13 pushed by copybara-service Bot
27m 32s test_979178228
test_979178228
27m 32s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulati… build #8117: Commit 5321954 pushed by copybara-service Bot
20m 0s dev
dev
20m 0s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulating the build #8116: Pull request #1024 synchronize by copybara-service Bot
20m 45s test_978577053
test_978577053
20m 45s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulati… build #8115: Commit 5321954 pushed by copybara-service Bot
21m 15s test_978577053
test_978577053
21m 15s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulating the build #8114: Pull request #1024 synchronize by copybara-service Bot
6m 34s test_978577053
test_978577053
6m 34s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulati… build #8113: Commit b1a69c6 pushed by copybara-service Bot
6m 42s test_978577053
test_978577053
6m 42s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulating the build #8112: Pull request #1024 opened by copybara-service Bot
26m 3s test_978577053
test_978577053
26m 3s
Decompose Gelu and FastGelu into GeluCdf and FastGeluCdf, reformulati… build #8111: Commit 05a5b90 pushed by copybara-service Bot
25m 43s test_978577053
test_978577053
25m 43s
Quantize softmax probabilities S into [0, 255] uint8_t and utilize un… build #8110: Commit 824b1b8 pushed by copybara-service Bot
25m 55s dev
dev
25m 55s