# Full GLM-5.3 on Four Sparks: Cache Layout, Speed and Fidelity Audit the 753B non-Flash EXL3 TP4 recipe, replicated and shared Q5 cache, million-token retrieval, BF16-reference fidelity and later RDMA fixes. - Type: Performance report - Evidence: source_reported; site reproduced: false - Model family: GLM - Model: bertholomus/GLM-5.3-EXL3-3.0bpw - Original author: bertholomus - Original project: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10 - Pinned document: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md - Source revision: 9a5c1100c68cd8ea832803c120ee63313049fc21 - Source commit time (not experiment date): 2026-10-09T05:31:45Z - Checked: 2026-10-18 - Experiment date: 2026-10-04 - Runtime topics: TensorFold - Hardware topics: spark, nvidia - Reported setup: {"system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "context_limit_tokens": 65536, "experiment_date": "2026-10-04"} - Performance figures: attributed selections with per-point conditions; no cross-source ranking ## Conditions Unknown is null. Recipe options are not measured benchmark settings. ```json { "conditions": { "model_revision": null, "weight_hash": null, "runtime_execution_revision": null, "input_tokens": null, "output_tokens": null, "concurrency": null, "cache_state": null, "quantization": null, "speculative_decoding": null, "temperature": null, "top_p": null }, "test_profiles": [ { "label": "replicated Q5 four code streams", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L57", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 4, "context_limit_tokens": 65536, "cache_state": "replicated", "metric_denominator": "source reports HTTP prompt time included; release concurrent raw usage/harness not retained", "token_accounting_verified": false, "available_serial_stream_probe_counts_nonempty_SSE_chunks": true, "release_was_computed_by_that_probe": null, "value_is_approximate": false } }, { "label": "DCP4 shared Q5 four code streams", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L68", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 4, "context_limit_tokens": 1048576, "cache_state": "four streams share one DCP4 pool", "metric_denominator": "source reports HTTP prompt time included; release concurrent raw usage/harness not retained", "token_accounting_verified": false, "available_serial_stream_probe_counts_nonempty_SSE_chunks": true, "release_was_computed_by_that_probe": null, "value_is_approximate": true } }, { "label": "32K-labelled prefill", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L71", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 1, "input_length_source_label": "32K", "context_limit_tokens": 1048576, "cache_state": null, "metric_denominator": "source HTTP prefill measurement, exact prompt count unreported" } }, { "label": "forward KL", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L167", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "forward KL" } }, { "label": "reverse KL", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L168", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "reverse KL" } }, { "label": "top-1 agreement", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L169", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "top-1 agreement" } }, { "label": "quant perplexity", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L170", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "quant perplexity" } }, { "label": "BF16 perplexity", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L170", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "BF16 perplexity" } }, { "label": "served engine prompt-path KL", "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L173", "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "served path; panel-specific cache details not retained", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "served engine" } } ], "recipe_parameters": null, "deployment_options": [] } ``` ## Reading notes - Selected values are source excerpts, not independent experiment counts. ## Limits - No independent Run AI Here model test. - Read source date, exact execution identity, denominator, memory scope and retries per observation. - No unmatched quality ranking or hardware speed ranking. ## License declaration {"declared_spdx": null, "scope": "Attribution and finite original factual interpretation only; models and dependencies retain their own terms."} ## Editorial answer The source runs full GLM-5.3 on four GB10 nodes with EXL3 and native MTP-3. Its two cache layouts trade reported aggregate code throughput for a larger shared window. Peak RAM and a task-accuracy suite are absent; the published quality panel measures quantization fidelity. By Run AI Here editorial ยท 2026-10-18 Local Spark Qwen draft; Codex source audit and rewrite. No site model benchmark. ## What must the deployment provide? Four 128 GB GB10 nodes use TP4, direct dual-port 200G RoCE, TensorFold 0.6.0 plus the model branch, EXL3 average 3.04 bpw and native MTP-3. It is text-only and does not use DFlash. The approximately 69.86 GiB per-rank weight fit is shape-derived and excludes whole-run cache and working memory. Sources: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md ## Which release produced the speed table? October 4 release numbers are attributed to engine prefix 3eb35dd; the later head ddfad35 adds per-device RDMA completion matching and graph prewarm. The full executed commit and checkpoint revision are unreported. Replicated Q5 at 65536/stream reports 82 aggregate code tokens/s for four streams; shared DCP4 Q5 at a 1048576-token pool reports about 64. Sources: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md ## How are HTTP accounting and fidelity limited? The README calls the rates HTTP measurements with prompt time included. Its available serial streaming probe counts nonempty SSE chunks rather than verified tokenizer/usage tokens, while the exact concurrent release harness/raw accounting is not retained here. The fidelity panel uses BF16 activations and no cache quantization, with forward KL 0.109, reverse KL 0.123 and 90.2 percent top-1 agreement. It is not a task-accuracy test. Sources: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md, https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_bench.py, https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_greedy.py ## What does the long-window result establish? The source finds a needle at 1039064 actual input tokens, reports 402 tokens/s at nominal 32K and about 76s first-token latency near 31K. Those do not demonstrate four full-million-token requests running together. Repetition counts, long-input peak RAM and raw per-request release accounting remain missing. Sources: https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md, https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_needle.py ## Selected source-reported figures Each point retains its own configuration, measurement unit and source. No site reproduction or cross-source ranking. ```json [ { "label": "replicated Q5 four code streams", "metric": "aggregate_output_rate", "value": 82, "unit": "tokens/s", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 4, "context_limit_tokens": 65536, "cache_state": "replicated", "metric_denominator": "source reports HTTP prompt time included; release concurrent raw usage/harness not retained", "token_accounting_verified": false, "available_serial_stream_probe_counts_nonempty_SSE_chunks": true, "release_was_computed_by_that_probe": null, "value_is_approximate": false }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 57, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L57" }, { "label": "DCP4 shared Q5 four code streams", "metric": "aggregate_output_rate", "value": 64, "unit": "tokens/s", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 4, "context_limit_tokens": 1048576, "cache_state": "four streams share one DCP4 pool", "metric_denominator": "source reports HTTP prompt time included; release concurrent raw usage/harness not retained", "token_accounting_verified": false, "available_serial_stream_probe_counts_nonempty_SSE_chunks": true, "release_was_computed_by_that_probe": null, "value_is_approximate": true }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 68, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L68" }, { "label": "32K-labelled prefill", "metric": "prefill_rate", "value": 402, "unit": "tokens/s", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "Q5 MLA latent cache; FP8 indexer", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": "2026-10-04", "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": 1, "input_length_source_label": "32K", "context_limit_tokens": 1048576, "cache_state": null, "metric_denominator": "source HTTP prefill measurement, exact prompt count unreported" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 71, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L71" }, { "label": "forward KL", "metric": "KL_reference_to_quant", "value": 0.109, "unit": "nats", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "forward KL" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 167, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L167" }, { "label": "reverse KL", "metric": "KL_quant_to_reference", "value": 0.123, "unit": "nats", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "reverse KL" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 168, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L168" }, { "label": "top-1 agreement", "metric": "top1_agreement", "value": 90.2, "unit": "percent", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "top-1 agreement" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 169, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L169" }, { "label": "quant perplexity", "metric": "perplexity", "value": 2.911, "unit": "perplexity", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "quant perplexity" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 170, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L170" }, { "label": "BF16 perplexity", "metric": "perplexity", "value": 2.75, "unit": "perplexity", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "no cache quantization", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "BF16 perplexity" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 170, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L170" }, { "label": "served engine prompt-path KL", "metric": "KL_reference_to_engine", "value": 0.111, "unit": "nats", "statistic": "source-reported observation; dispersion unreported", "uncertainty": null, "conditions": { "model": "GLM-5.3 (full 753B, not Flash)", "model_artifact": "bertholomus/GLM-5.3-EXL3-3.0bpw", "model_revision": null, "engine": "TensorFold 0.6.0 + glm-dsa-tp4 branch", "runtime_execution_revision": null, "measured_runtime_short": "3eb35dd", "later_branch_head_short": "ddfad35", "executed_image_digest": null, "system": "four DGX Spark GB10, 128 GB unified memory each, direct dual-port 200G RoCE", "node_count": 4, "parallelism": "TP4", "weight_precision": "EXL3 3.04 bpw whole-model average; exllamav3 1.5.3", "kv_dtype": "served path; panel-specific cache details not retained", "draft": "native MTP-3", "vision": false, "DFlash": false, "experiment_date": null, "temperature": 0, "thinking": null, "input_tokens": null, "output_tokens": null, "host_peak_RAM": null, "repetitions": null, "concurrency": null, "quality_panel": "wikitext-2 test, 32 x 2048 = 65536 positions", "activation_dtype": "BF16", "reference": "BF16 full model", "cache_state": null, "quality_is_task_accuracy": false, "quality_arm": "served engine" }, "evidence_state": "source_reported", "site_reproduced": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 173, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L173" } ] ``` ## Supplemental source identities ```json { "supplemental_sources": [ { "document_id": "glmfull-speed-method", "parent_record_id": "glm53-full-tp4-exl3", "publisher": "bertholomus", "repository": "bertholomus/glm-5.3-tensorfold-tp4-4xgb10", "revision": "9a5c1100c68cd8ea832803c120ee63313049fc21", "document_path": "bench/tf_bench.py", "pinned_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_bench.py", "document_sha256": "9652bf8a7359414ad451dc05ceee4f4bef6cfcf489e65334e9bec1e7895192b2", "last_commit_at": "2026-10-09T05:31:45Z", "checked_on": "2026-10-18", "public_URL_is_immutable": true, "revision_kind": "Git source document revision; executed engine and weight revisions are separate fields" }, { "document_id": "glmfull-needle-method", "parent_record_id": "glm53-full-tp4-exl3", "publisher": "bertholomus", "repository": "bertholomus/glm-5.3-tensorfold-tp4-4xgb10", "revision": "9a5c1100c68cd8ea832803c120ee63313049fc21", "document_path": "bench/tf_needle.py", "pinned_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_needle.py", "document_sha256": "cdf8ab15fc7ec88073a714696b561f2e7a292ac9a0490571935bbe2cc933a4ab", "last_commit_at": "2026-10-09T05:31:45Z", "checked_on": "2026-10-18", "public_URL_is_immutable": true, "revision_kind": "Git source document revision; executed engine and weight revisions are separate fields" }, { "document_id": "glmfull-greedy-method", "parent_record_id": "glm53-full-tp4-exl3", "publisher": "bertholomus", "repository": "bertholomus/glm-5.3-tensorfold-tp4-4xgb10", "revision": "9a5c1100c68cd8ea832803c120ee63313049fc21", "document_path": "bench/tf_greedy.py", "pinned_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/bench/tf_greedy.py", "document_sha256": "834c9f1e73af9431b585adf00aa3f14254d8713b638a4fa95c36b6a2074dbf37", "last_commit_at": "2026-10-09T05:31:45Z", "checked_on": "2026-10-18", "public_URL_is_immutable": true, "revision_kind": "Git source document revision; executed engine and weight revisions are separate fields" } ] } ``` ## Paired source quality comparisons ```json { "quality_comparisons": [ { "panel": "32x2048 wikitext-2", "BF16_activations": true, "cache_quantization": false, "forward_KL": 0.109, "reverse_KL": 0.123, "top1_percent": 90.2, "quant_PPL": 2.911, "BF16_PPL": 2.75, "task_accuracy": null } ] } ``` ## memory observations ```json { "memory_observations": [ { "kind": "shape-derived weights per rank, not peak", "value": 69.86, "unit": "GiB", "is_measured_peak": false, "whole_system_peak": null, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 121, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L121" } ] } ``` ## long context cases ```json { "long_context_cases": [ { "actual_input_tokens": 1039064, "needle_found": true, "independent_repeats": null, "concurrent_full_window_requests_proven": false, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 71, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L71" } ] } ``` ## case reports ```json { "case_reports": [ { "classification": "source_reported_RDMA_completion_collision", "old_error": "ibv_post_recv failed", "fix_short": "cad0d2a", "measured_release_short": "3eb35dd", "later_head_short": "ddfad35", "full_execution_revision": null, "source_document_id": "glmfull-tp4-readme", "source_document_sha256": "088db2d55fbb6aecc2cc6576b3db9a26e70a13a2b5d89b83ad1cb0b54b03a6b4", "source_line": 35, "evidence_url": "https://github.com/bertholomus/glm-5.3-tensorfold-tp4-4xgb10/blob/9a5c1100c68cd8ea832803c120ee63313049fc21/README.md#L35" } ] } ``` ## quality comparisons ```json { "quality_comparisons": [ { "panel": "32x2048 wikitext-2", "BF16_activations": true, "cache_quantization": false, "forward_KL": 0.109, "reverse_KL": 0.123, "top1_percent": 90.2, "quant_PPL": 2.911, "BF16_PPL": 2.75, "task_accuracy": null } ] } ```